IJACSA Volume2No11

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011

(i)
www.ijacsa.thesai.org
IJACSA Editorial
From the Desk of Managing Editor
It is a pleasure to present our readers with the November 2011 Issue of International Journal of Advanced Computer Science
and Applications (IJACSA).
The renaissance stimulated by the field of Computer Science is generating multiple formats and channels of communication and
creativity. IJACSA is one of the most prominent publications in the field and engaging the ubiquitous spread of subject
knowledge with effectiveness in all classes of audience. Nevertheless, the promise of increased engagement requires that we
consider how this might be accomplished, delivering up-to-date and authoritative coverage of advanced computer science and
applications.
The journal has a wide scope ranging from the many facets of methodological foundations to the details of technical issues and
the aspects of industrial practice. It includes articles related to research findings, technical evaluations, and reviews. In addition
it provides a forum for the exchange of information on all aspects.
The editorial board of the IJACSA consists of individuals who are committed to the search for high-quality research suitable for
publication. These individuals, working with the editor to achieve IJACSA objectives, assess the quality, relevance, and
readability of individual articles.
The contents include original research and innovative applications from all parts of the world. This interdisciplinary journal has
brought together researchers from academia and industry as well as practitioners to share ideas, problems and solutions
relating to computer science and application with its convergence strategies, and to disseminate the most innovative research.
As a consequence only 26% of the received articles have been finally accepted for publication.
Therefore, IJACSA in general, could serve as a reliable resource for everybody loosely or tightly attached to this field of science.
The published papers are expected to present results of significant value to solve the various problems with application services
and other problems which are within the scope of IJACSA. In addition, we expect they will trigger further related research and
technological improvements relevant to our future lives.
We hope to continue exploring the always diverse and often astonishing fields in Advanced Computer Science and Applications.
Thank You for Sharing Wisdom!

Managing Editor
IJACSA
Volume 2 Issue 11, November 2011
editorijacsa@thesai.org
ISSN 2156-5570 (Online)
ISSN 2158-107X (Print)
2011 The Science and Information (SAI) Organization
Vol. 2, No. 11, 2011
(ii)
IJACSA Associate Editors
Dr. Zuqing Zhu
Service Provider Technology Group of Cisco Systems, San Jose
Domain of Research: Research and development of wideband access routers for hybrid fibre-
coaxial (HFC) cable networks and passive optical networks (PON)
Dr. Ka Lok Man
Department of Computer Science and Software Engineering at the Xi'an Jiaotong-Liverpool
University, China
Domain of Research: Design, analysis and tools for integrated circuits and systems; formal
methods; process algebras; real-time, hybrid systems and physical cyber systems;
communication and wireless sensor networks.
Dr. Sasan Adibi
Technical Staff Member of Advanced Research, Research In Motion (RIM), Canada
Domain of Research: Security of wireless systems, Quality of Service (QoS), Ad-Hoc Networks,
e-Health and m-Health (Mobile Health)
Dr. Sikha Bagui
Associate Professor in the Department of Computer Science at the University of West Florida,
Domain of Research: Database and Data Mining.
Dr. T. V. Prasad
Dean, Lingaya's University, India
Domain of Research: Bioinformatics, Natural Language Processing, Image Processing, Expert
Systems, Robotics
Dr. Bremananth R
Research Fellow, Nanyang Technological University, Singapore
Domain of Research: Acoustic Holography, Pattern Recognition, Computer Vision, Image
Processing, Biometrics, Multimedia and Soft Computing
Vol. 2, No. 11, 2011
(iii)
IJACSA Reviewer Board

A Kathirvel
Karpaga Vinayaka College of Engineering and
Technology, India
Abbas Karimi
I.A.U_Arak Branch (Faculty Member) & Universiti
Putra Malaysia
Dr. Abdul Wahid
Gautam Buddha University, India
Abdul Khader Jilani Saudagar
Al-Imam Muhammad Ibn Saud Islamic University
Abdur Rashid Khan
Gomal Unversity
Dr. Ahmed Nabih Zaki Rashed
Menoufia University, Egypt
Ahmed Sabah AL-Jumaili
Ahlia University
Md. Akbar Hossain
Aalborg University, Denmark and AIT, Greeceas
Albert Alexander
Kongu Engineering College,India
Prof. Alcinia Zita Sampaio
Technical University of Lisbon
Amit Verma
Rayat & Bahra Engineering College, India
Ammar Mohammed Ammar
Department of Computer Science, University of
Koblenz-Landau
Arash Habibi Lashakri
University Technology Malaysia (UTM), Malaysia
Asoke Nath
St. Xaviers College, India
B R SARATH KUMAR
Lenora College of Engineering, India
Binod Kumar
Lakshmi Narayan College of Technology, India
Bremananth Ramachandran
School of EEE, Nanyang Technological University
Dr.C.Suresh Gnana Dhas
Park College of Engineering and Technology, India
Mr. Chakresh kumar
Manav Rachna International University, India
Chandra Mouli P.V.S.S.R
VIT University, India
Chandrashekhar Meshram
Shri Shankaracharya Engineering College, India
Constantin POPESCU
Department of Mathematics and Computer
Science, University of Oradea
Prof. D. S. R. Murthy
SNIST, India.
Deepak Garg
Thapar University.
Prof. Dhananjay R.Kalbande
Sardar Patel Institute of Technology, India
Dhirendra Mishra
SVKM's NMIMS University, India
Divya Prakash Shrivastava
EL JABAL AL GARBI UNIVERSITY, ZAWIA
Dragana Becejski-Vujaklija
University of Belgrade, Faculty of organizational
sciences
Fokrul Alom Mazarbhuiya
King Khalid University
G. Sreedhar
Rashtriya Sanskrit University
Ghalem Belalem
University of Oran (Es Senia)
Hanumanthappa.J
University of Mangalore, India
Dr. Himanshu Aggarwal
Punjabi University, India
Huda K. AL-Jobori
Ahlia University
Dr. Jamaiah Haji Yahaya
Northern University of Malaysia (UUM), Malaysia
Jasvir Singh
Communication Signal Processing Research Lab
Jatinderkumar R. Saini
S.P.College of Engineering, Gujarat
Prof. Joe-Sam Chou
Nanhua University, Taiwan
Dr. Juan Jos Martnez Castillo
Yacambu University, Venezuela
Dr. Jui-Pin Yang
Vol. 2, No. 11, 2011
(iv)
Shih Chien University, Taiwan
Dr. K.PRASADH
Mets School of Engineering, India
Ka Lok Man
Xian Jiaotong-Liverpool University (XJTLU)
Dr. Kamal Shah
St. Francis Institute of Technology, India
Kodge B. G.
S. V. College, India
Kohei Arai
Saga University
Kunal Patel
Ingenuity Systems, USA
Lai Khin Wee
Technischen Universitt Ilmenau, Germany
Latha Parthiban
SSN College of Engineering, Kalavakkam
Mr. Lijian Sun
Chinese Academy of Surveying and Mapping, China
Long Chen
Qualcomm Incorporated
M.V.Raghavendra
Swathi Institute of Technology & Sciences, India.
Madjid Khalilian
Islamic Azad University
Mahesh Chandra
B.I.T, India
Mahmoud M. A. Abd Ellatif
Mansoura University
Manpreet Singh Manna
SLIET University, Govt. of India
Marcellin Julius NKENLIFACK
University of Dschang
Md. Masud Rana
Khunla University of Engineering & Technology,
Bangladesh
Md. Zia Ur Rahman
Narasaraopeta Engg. College, Narasaraopeta
Messaouda AZZOUZI
Ziane AChour University of Djelfa
Dr. Michael Watts
University of Adelaide, Australia
Miroslav Baca
University of Zagreb, Faculty of organization and
informatics / Center for biomet
Mohamed Ali Mahjoub
Preparatory Institute of Engineer of Monastir
Mohammad Talib
University of Botswana, Gaborone
Mohammed Ali Hussain
Sri Sai Madhavi Institute of Science & Technology
Mohd Helmy Abd Wahab
Universiti Tun Hussein Onn Malaysia
Mohd Nazri Ismail
University of Kuala Lumpur (UniKL)
Mueen Uddin
Universiti Teknologi Malaysia UTM
Dr. Murugesan N
Government Arts College (Autonomous), India
Nitin S. Choubey
Mukesh Patel School of Technology Management &
Eng
Dr. Nitin Surajkishor
NMIMS, India
Paresh V Virparia
Sardar Patel University
Dr. Poonam Garg
Institute of Management Technology, Ghaziabad
Raj Gaurang Tiwari
AZAD Institute of Engineering and Technology
Rajesh Kumar
National University of Singapore
Rajesh K Shukla
Sagar Institute of Research & Technology-
Excellence, India
Dr. Rajiv Dharaskar
GH Raisoni College of Engineering, India
Prof. Rakesh. L
Vijetha Institute of Technology, India
Prof. Rashid Sheikh
Acropolis Institute of Technology and Research,
India
Ravi Prakash
University of Mumbai
Rongrong Ji
Columbia University
Dr. Ruchika Malhotra
Delhi Technological University, India
Dr.Sagarmay Deb
University Lecturer, Central Queensland University,
Australia
Vol. 2, No. 11, 2011
(v)
Saleh Ali K. AlOmari
Universiti Sains Malaysia
Dr. Sana'a Wafa Al-Sayegh
University College of Applied Sciences UCAS-
Palestine
Santosh Kumar
Graphic Era University, India
Sasan Adibi
Research In Motion (RIM)
Saurabh Pal
VBS Purvanchal University, Jaunpur
Seyed Hamidreza Mohades Kasaei
University of Isfahan
Shahanawaj Ahamad
The University of Al-Kharj
Shaidah Jusoh
University of West Florida
Sikha Bagui
Zarqa University
Dr. Smita Rajpal
ITM University
Suhas J Manangi
Microsoft
SUKUMAR SENTHILKUMAR
Universiti Sains Malaysia
Sunil Taneja
Smt. Aruna Asaf Ali Government Post Graduate
College, India
Dr. Suresh Sankaranarayanan
University of West Indies, Kingston, Jamaica
T C.Manjunath
Visvesvaraya Tech. University
T V Narayana Rao
Hyderabad Institute of Technology and
Management, India
T. V. Prasad
Lingaya's University
Taiwo Ayodele
Lingaya's University
Totok R. Biyanto
Infonetmedia/University of Portsmouth
Varun Kumar
Institute of Technology and Management, India
Vellanki Uma Kanta Sastry
Sreeneedhi
Dr. V. U. K. Sastry
SreeNidhi Institute of Science and Technology
(SNIST), Hyderabad, India.
Vinayak Bairagi
Sinhgad Academy of engineering, India
Vitus S.W. Lam
The University of Hong Kong
Vuda Sreenivasarao
St.Marys college of Engineering & Technology,
Hyderabad, India
Y Srinivas
GITAM University
Mr.Zhao Zhang
City University of Hong Kong, Kowloon, Hong Kong
Zhixin Chen
ILX Lightwave Corporation
Zuqing Zhu
University of Science and Technology of China
Vol. 2, No. 11, 2011
(vi)
CONTENTS

Paper 1: Crytosystem for Computer security using Iris patterns and Hetro correlators
Authors: R. Bremananth, Ahmad Sharieh
PAGE 1 8

Paper 2: On the transmission capacity of quantum networks
Authors: Sandra Knig, Stefan Rass

PAGE 9 16

Paper 3: Confidential Deterministic Quantum Communication Using Three Quantum States
Authors: Piotr ZAWADZKI
PAGE 17 20

Paper 4: A Novel Implementation of RISI Controller Employing Adaptive Clock Gating Technique
Authors: M.Kamaraju, Praveen V N Desu
PAGE 21 27

Paper 5: Strength of Quick Response Barcodes and Design of Secure Data Sharing System
Authors: Sona Kaushik
PAGE 28 32

Paper 6: Graphing emotional patterns by dilation of the iris in video sequences
Authors: Rodolfo Romero Herrera, Francisco Gallegos Funes, Saul De La O Torres
PAGE 33 37

Paper 7: The impact of competitive intelligence on products and services innovation in organizations
Authors: Phathutshedzo Nemutanzhela, Tiko Iyamu
PAGE 38 44

Paper 8: Arabic Sign Language (ArSL) Recognition System Using HMM
Authors: Aliaa A. A.Youssif , Amal Elsayed Aboutabl, Heba Hamdy Ali
PAGE 45 51

Paper 9: Modularity Index Metrics for Java-Based Open Source Software Projects
Authors: Andi Wahju Rahardjo Emanuel, Retantyo Wardoyo, Jazi Eko Istiyanto, Khabib Mustofa
PAGE 52 58

Paper 10: Survey of Nearest Neighbor Condensing Techniques
Authors: MILOUD-AOUIDATE Amal, BABA-ALI Ahmed Riadh
PAGE 59 64
Vol. 2, No. 11, 2011
(vii)

Paper 11: Concurrent Edge Prevision and Rear Edge Pruning Approach for Frequent Closed Itemset
Mining
Authors: Anurag Choubey, Dr. Ravindra Patel, Dr. J.L. Rana
PAGE 65 70

Paper 12: Error Filtering Schemes for Color Images in Visual Cryptography
Authors: Shiny Malar F.R, Jeya Kumar M.K
PAGE 71 76

Paper 13: Passwords Selected by Hospital Employees: An Investigative Study
Authors: B. Dawn Medlin, Ken Corley, B. Adriana Romaniello
PAGE 77 81

Paper 14: Current Trends in Group Key Management
Authors: R. Siva Ranjani, Dr.D.Lalitha Bhaskari, Dr.P.S.Avadhani
PAGE 82 86

Paper 15: CluSandra: A Framework and Algorithm for Data Stream Cluster Analysis
Authors: Jose R. Fernandez, Eman M. El-Sheikh
PAGE 87 99

Paper 16: Clustering: Applied to Data Structuring and Retrieval
Authors: Ogechukwu N. Iloanusi, Charles C. Osuagwu
PAGE 100 105

Paper 17: Irrigation Fuzzy Controller Reduce Tomato Cracking
Authors: Federico Hahn
PAGE 106 109

Paper 18: Plethora of Cyber Forensics
Authors: N.Sridhar, Dr.D.Lalitha Bhaskari, Dr.P.S.Avadhani
PAGE 110 114

Paper 19: A Fuzzy Similarity Based Concept Mining Model for Text Classification
Authors: Shalini Puri
PAGE 115 121

Paper 20: Improved Echo cancellation in VOIP
Authors: Patrashiya Magdolina Halder, A.K.M. Fazlul Haque
PAGE 122 125

Vol. 2, No. 11, 2011
(viii)
Paper 21: A New Test Method on the Convergence and Divergence for Infinite Integral
Authors: Guocheng Li
PAGE 126 129

Paper 22: Adaptive Outlier-tolerant Exponential Smoothing Prediction Algorithms with Applications to
Predict the Temperature in Spacecraft
Authors: Hu Shaolin, Zhang Wei, Li Ye, Fan Shunxi
PAGE 130 133

Paper 23: Towards Quranic reader controlled by speech
Authors: Yacine Yekache, Yekhlef Mekelleche, Belkacem Kouninef
PAGE 134 137

Paper 24: Design and Implementation for Multi-Level Cell Flash Memory Storage Systems
Authors: Amarnath Gaini, K Vijayalaxmi, Sathish Mothe
PAGE 138 143

Paper 25: A novel approach for pre-processing of face detection system based on HSV color space and
IWPT
Authors: Megha Gupta, Prof. Neetesh Gupta
PAGE 144 147

Paper 26: A new approach of designing Multi-Agent Systems
Authors: Sara Maalal, Malika Addou
PAGE 148 157
Vol. 2, No. 11, 2011
1 | P a g e
Crytosystem for Computer security using Iris patterns
and Hetro correlators

R. Bremananth
Information Systems and Technology Department,
Sur University College,
Sur, Oman.
Ahmad Sharieh
Information Systems & Technology Department,
Sur University College,
Sur, Oman.

AbstractBiometric based cryptography system provides an
efficient and secure data transmission as compare to the
traditional encryption system. However, it is a computationally
challenge task to solve the issues to incorporate biometric and
cryptography. In connection with our previous works, this paper
reveals a robust cryptosystem using iris biometric pattern as a
crypto-key to resolve the issues in the encryption. An error
correction engine based on hetro-correlators has been used to
evoke the partially tarnished data fashioned by the decryption
process. This process determines the non-repudiation and key
management problems. The experimental results show that the
suggestion algorithm can implement in the real-life cryptosystem.
Keywords-Auto-correlators; Biometric; crytosystem; Hetro-
correlators.
I. INTRODUCTION
Cryptography provides a secure proliferation of information
exchange across the insecure data communication [1]. It
authenticates messages based on the mathematical key but not
based on the real-life user those who are the genuine owner.
Traditional cryptosystem requires a lengthy key to encrypt and
decrypt in sending and receiving the messages, respectively.
But these keys can be guessed or cracked. Moreover,
maintaining and sharing lengthy, random keys in enciphering
and deciphering process is the critical problem in the
cryptography system. A new approach is described for
generating a crypto key, which is acquired from iris patterns. In
the biometric field, template created by the biometric algorithm
can only be authenticated with the same person. Among the
biometric templates, iris features can efficiently be
distinguished with individuals and produces less false positives
in a large population. This type of iris code distribution
provides merely less intra-class variability that aids the
cryptosystem to confidently decrypt messages with an exact
matching of iris pattern. In traditional cryptography system,
key management is a cumbersome process that is, key must be
generated each time with an extensive computational process
and the dissemination of keys is also a very difficult process at
the non-secure channels [1]. It consumes lot of system time and
produces overburden to the application domains. In addition,
non-repudiation cannot easily be handled in the traditional
cryptosystem.
The Biometric key cryptography (BKC) is an emerging
reliable alterative that can be used to resolve key management,
large key computational process and address the non-
repudiation problems [2]. In the cryptography system, data will
be secured using a symmetric cipher system and in public-key
system digital signatures are used for secure key exchange
between users. However, in both systems the dimension of
security accuracy is dependent on the cryptography strong
keys. They are required to remember and enter the large key
whenever needed. Instead of remembering large keys, the user
may opt to give password to encrypt and decrypt the
cryptography keys. There is no direct tie up between user and
password that is, the system running the cryptography
algorithm is unable to differentiate the genuine user and
impostors who are unauthorized to work with the system.
Thus, a reliable alternative to the password security is the
biometric guard for the cryptography keys. Whenever user
wishes to access through a secured key, biometric sample is
captured, authenticated by the classifiers and then key is
released to encipher / decipher the desired data. In general
biometric cryptosystem has been classified by three categories.
The first method is to release the cryptography key from secure
area in accordance with biometric matching algorithm. It
requires the secured communication line to avoid
eavesdroppers attacks. Furthermore, if the user may store the
biometric templates or crypto keys in workstation machines
then the system becomes an insecure one. In the next method,
the crypto key is embedded as a part of biometric template in a
specific location. However, if impostors may determine the
location of the keys, again it becomes catastrophic to the
system. The third method is based on using biometric features
as cryptography keys, which gives more secure manner of
proliferation of information exchange.
The proposed approach is broadly classified into three
phases. The first phase is related with compact way to obtain
iris feature codes from the human irises. The second one
describes the algorithm to encrypt and decrypt the messages
using iris bits. In the third phase, the error correction engine is
employed to recall the partially corrupted bits generated in the
decryption using associative memories. The issue of biometric
pattern is the partially varied features produced in the feature
extraction process, which subsequently makes partially
corrupted data in the decryption process. This dissimilarity may
occur due to environments, illuminations, distance variation
and other artifacts. However more stable pattern produced by
the iris is secured in the persons lifetime and produces limited
Vol. 2, No. 11, 2011
2 | P a g e
number of bits variations in the features, which assists to
decrypt the messages in massive manner. In addition, re-
enrolment of iris keys is required to preserve the system
security more consistently.
In the current literature several studies were proposed
related with biometric cryptosystem but most of them dealt
with fingerprints and few of them were concerned with iris
features. Albert Bodo proposed a method of directly using
biometric as cryptography key in the patent of German [1]. In
(Davida et al. [2][3]), 2048-bit iris code was used for
enciphering and deciphering process. Key generation is
invoked based on the error bits of the iris codes. This system
stored the error correction bits along with iris keys inside the
database. Thus, impostors may eavesdrop key information and
a count of error correction bits from the local database. In
(Linnartz et al. [4], Clancy et al. [5], Monrose et al. [6]), the
key generation was based on biometrics such as fingerprints
[18] and voices, but they required more calculations to release
the key than the traditional cryptography system. The problem
of generating cryptograph key from face biometric features had
been studied by Yao-Jen Chang et al. [7]. The survey of multi-
biometric cryptosystems was discussed by Uludag et al. [8]. A
method of iris compression for cryptography documentation on
off-line verification was proposed by Daniel et al. [9]. In this
study, a modified Fourier-Mellin transformation was employed
to create iris template for representing EyeCert system, which
consists of two components. The first one is details of personal
data related with the subjects, and the second one is the iris
feature encoded in the form of barcodes. In another study of iris
biometric cryptosystem, Feng Hao et al. [10] proposed a
method based on error-free iris key that was devised using a
two-layer error correction technique incorporated with
Hadamard and Reed-Solomon codes. The extracted code was
saved in a tamper-resistant token such as a smart card. In our
previous work, (Bremananth et al. [11]) proposed auto-
correlator to recoup the corrupted bio-metric crypto key. In this
paper, a robust hetro-correlator has been proposed to regain the
data.
The block diagram of the proposed iris cryptosystem is
illustrated in Fig. 1. It suggests a compact way to extract
feature from the iris patterns and these features are treated as
crypto key for the on-line cryptography system. This system
outperforms other traditional approaches and provides an
efficient solution for non-repudiation approach as well. It
employs 135-bit iris code which is extracted by wavelet
analysis[12][13][14] and applying these codes in enciphering
and deciphering of the input stream of binary data which might
be originating from voice, text, video, image or other sources.
Next, the auto-correlators and hetero-correlators are used to
recall original bits from the partially corrupted data produced in
the decryption process. It intends to resolve the repudiation and
key management problems. However, the performance of error
correction model dependents on the correlators used in the
system. Hence the guarantee issues of these methods were
verified and the experimental results were analyzed in both
symmetric iris cryptosystem (SIC) and non-repudiation iris
cryptosystem (NRIC). It shows that this new approach provides
considerably high authentication in enciphering and
deciphering processes. The remainder of the paper has been
organized as follows. Section II describes the symmetric iris
cryptosystem. Non-repudiation cryptosystem is described in
Section III. Error correction engines and their functionalities
are given in Section IV. Section V describes the experimental
results of the bio-metric cryptosystem and concluding remarks
are given in Section VI.

Figure 1. A proposed block diagram of the iris cryptography system.
II. SYMMETRIC IRIS CRYPTOSYSTEM
Iris patterns are used for fabricating a key to encipher and
decipher the plain text in between sender and receiver over
insecure channels [2][11]. The advantages of iris cryptosystem
are to reduce the system processing time to make a complex
key for standard cryptography algorithm and to generate cipher
keys without getting back from complex key generation
sequences. The identical iris code is used in both ends to
encrypt and decrypt the message in the SIC system. In order to
decrypt a message, the recipient needs an identical copy of the
iris code. Figure 2 shows the iris based symmetric
cryptography system. The transmission of enrolled iris code
over the channel is vulnerable to eavesdropping. Hence, the
copy of the enrolled iris code is needed in the recipient side,
which is being used by the decryption process. In this
approach, XOR operation is used to encrypt and decrypt the
message. The significant steps of SIC encryption algorithm is
described as follows:
Step 1: Let K be the key sequence
p 2 1
I ,..., I , I produced by iris feature encoding algorithm
for the encryption transformation. In the experiment 136-bit
key sequence (135-bit iris code and one padding bit) is used in
the encryption process.
Step 2: Let S be a source alphabet of N
symbols
N
S ,..., S , S
2 1
. Each alphabet in S is converted to its
equivalent 8-bit binary strings. The bits of messages undergo
XORing with iris key sequence and generate a non-breakable
cipher-bit described as
) I ,..., I , I S ,..., S , S ( Ency C
p N i 2 1 2 1
= (1)
Vol. 2, No. 11, 2011
3 | P a g e
where
i
C is set of cipher bits. The decryption algorithm is
described as follows:
Step 1: The testing iris pattern is extracted and iris codes
are formed. The iris-matching algorithm verifies the test and
the enrol iris codes. If weighted distance (WD) is
19 . 0 WD 0 s s , then the matched enrolled iris code is used
for deciphering the messages, otherwise rejected.
Step 2: Let
K
p
I ,...,
2
I ,
1
I e
be an enrolled iris code and
n
C C C ,...,
2
,
1
is a set of cipher text produced by the encryption
process. Enrolled iris codes are XORed with set of cipher bits
and generate the original messages using Equation (2).
Bit
Stream
Iris
Encoding
Iris key 136-bit
+
Encryption
Encrypted Bit Sequences
Decryption
Iris Matching algorithm
+
Is genuine
Iris Key 136-bit
Yes
No
Reject
Iris code
(On-line)
Original Bit
stream

Figure 2. The process of SIC system.
) I ,..., I , I C ,..., C , C ( Decy S
p n i 2 1 2 1
= (2)
where
i
S is set of source alphabet bits and I = 1,2, 3, N.
In the SIC system, key dissemination problem is completely
avoided. However, the system needs iris database and iris-
matching algorithm in the decryption process to get back the
original messages. In order to resolve repudiation problem, the
iris database and iris-matching algorithm are eliminated from
the SIC system. The detailed description of this process is
discussed in the next section.
III. NON-REPUDIATION IRIS CRYPTOSYSTEM

Unlike SIC system, the NRIC system bypasses the iris-
matching process and do not access iris database in the
decryption process. The testing iris code can directly be
XORed with cipher bits transmitted by the encryption process
as illustrated in Fig. 3. Iris codes are changed from session to
session with minimum variation (WD<=0.19) for the same
subject eye. Hence the decryption process may produce the
probability of partially corrupted cipher bits ranging from 0 to
0.19. Perhaps, if intruder may tap the cipher bits at the non-
secure channels then the probability of decrypting the message
is complicated from 0.2 to 1 partially corrupted bit in every
135-bit iris code. Thus, it produces more complexity to the
intruder to get back the original messages. But the cipher bits
accessed by the genuine subjects have probability of error rate
at most 0.19, so that, less complexity have been created in the
decryption process.
Bit
Stream
Iris
Encoding
Iris key 136-bit
+
Encryption
Encrypted Bit Sequences
Decryption
Iris Encoding (On-
line)
Original Bit
stream
Iris Key 136-bit
Error bits correction using
auto/Hetero correlators
Partially corrupted bit
sequences
+

Figure 3. A sequence sketch of Non-repudiation cryptosystem.
In this method cipher bits are directly XORed with the test
iris key and produce the partially corrupted bits. These are very
close to the original message if the test iris key is actually
extracted from the genuine subject; otherwise the partially
corrupted bits are larger than the threshold maintained in the
system.
Thus impostors can be restricted to access the original
scripts. The error bit correction module subsequently corrects
these bits by using the two different correction engines such as
either auto-correlators or hetero-correlators that perform the
probability of error correction based on iris-weighted distance.
Thus this process overcomes repudiation problem and reduces
the key management issues. However, the performance of the
NRIC fully depends on the guarantee of the error correction
engines because recalling the original bits is a difficult process
in the real time processing of encryption and decryption.
IV. ERROR CORRECTION ENGINES
In the process of biometric cryptosystem, the major
limitation is a way to get back the original bits from the
partially corrupted bits generated by the decryption. In the
literature, several studies had been performed to recall the
trained patterns from the partially corrupted patterns. Bart
Kosko et al. [15] enhanced the bidirectional associative
memories (BAM), which behaves as a hetero-associative
content addressable memory (CAM) storing and recalling the
vector pairs.
The bidirectional associative memory with multiple training
can be guaranteed to recall a single trained pair under suitable
initial conditions of data. Sufficient condition for a correlation
matrix to make the energies of the training pairs was described
by Yeou-Fang et al. [16]. An essential condition for
generalization of correlation matrix of BAM which guarantees
the recall of all the training pairs was discussed by Yeou-Fang
et al. [17]. This paper adopts two different methods to recall the
corrupted patterns. The first one is related to auto-associative
and the other one is concerned with hetero-associative.
Vol. 2, No. 11, 2011
4 | P a g e
A. Autocorrelators
Associative memories are one of the key models of
neural network and they can act as a human brain to recall the
associated patterns perfectly from the corrupted patterns. If the
associated pair (x, y) is the identical pattern, then the model of
associative memory is called as auto-associative memory. For
the recall operation, auto-associatives require the correlation
memory or connection matrix, which aids to retrieve original
patterns from the partially corrupted pattern. It is called as auto-
correlators and is adopted in the error correction process of
NRIC. The algorithm of error bits correction process is
described as follows [11]:
Step 1: The partially corrupted data obtained in the
decryption process is taken for further processing. This data is
transformed to bipolar patterns (
c
|
). Let M be the number of
stored bipolar patterns
m
p p p ,...,
2
,
1
and i
th
patterns is
) (
, , 2
,
1 in i i
p p p
A
where n is the number of bits in the stored
pattern. The connection matrix CM is derived as
| | | | n j for n i for p p CM
i
n
i
T
i ij
.. 1 , ... 1
1
= = =
=
(3)
Step 2: The auto-correlator recalls the original patterns (u)
using
m j for p CM g
j cj j
.. 1 ) ), (( = - = | u
(4)
<
=
>
=
0 1
0
0 1
) , (
_
_
_
_
if
if
if
g
(5)
where u
j
is the recalled original pattern,
c
| is a partially
corrupted data and ) , ( _ g is the threshold function.
Step 3: Repeat Step 2 until
u | >
=
n
i
i i
1
, where is a
vigilance parameter.
The parameter provides minimum error bit correction in
between the genuine subject iris code and partially corrupted
cipher bits. This parameter gives more complexity to the
intruder to get back the original messages. For example, if the
patterns are | | 1 1 1
1
= p , | | 1 1 1
2
= p ,
| | 1 1 1
3
= p then the connection matrix (CM)
is:
(
(
(
3 3 1
3 3 1
1 1 3

If partially corrupted data produced in the decryption
process is | | 1 1 1 = p then the computation with CM
produce the threshold conditions: g(-3,-1),g(-1,1) and g(1,1). It
gives the original pattern | | 1 1 1 = O .
B. Heterocorrelators
In this approach, noisy variation of different types of iris
codes are not explicitly estimated and stored in the verification
database [17]. If they may explicitly be estimated, then it leads
to leak of security information to the adversary. Hence, hetero-
correlations are directly used to recall the original patterns from
the corrupted patterns that need not have any additional
information such as noisy variations. This is nothing but an
associative memory, which is an imitation model of human
brains ability to recall associate patterns. In the non-
repudiation cryptosystem, the decryption produces noise bits
which should be corrected properly and converted to its real bit
sequences. If the associated pattern pairs (x, y) are different,
then this model recalls y. If x is given, then y can be called.
This is referred as hetero-associative memory. This memory is
used to recall the original patterns from the corrupted patterns.
For the recall operation, hetero-associative requires a
correlation memory or connection matrix, which aids to
retrieve original patterns. This is so-called hetero-correlators.
The algorithm of error bits correction process is described as
follows:
Step 1: The partially corrupted data obtained in the
decryption process is taken for further processing. This data is
transformed into its bipolar patterns (o ). Let M be the number
of stored bipolar pairs given as
} , { },...,
2
,
2
{ },
1
,
1
{
m
Q
m
P Q P Q P
(6)
where
} ,... { }, ,... {
2 , 1 2 , 1 io i i i in i i i
q q q Q p p p P = =
, P and Q
represent stored and exemplar patterns of distorted bipolar data,
respectively. The connection matrix (CM) is derived as

| | | | o .. j for , n ... i for Q P CM
i
n
i
T
i i ij
1 1
1
= = =
=
E
(7)
Vol. 2, No. 11, 2011
5 | P a g e
where CM is a correction matrix used in the hetero-
correlation process and E is a set of energy constants i.e.,
+
eR E , R is a set of real numbers. Calculate
'
k and
k from Equations (8) and (9) and assign to
'
o and o ,
respectively.
Step 2: The hetero-correlator recalls the original bit
sequences ( ) using
) (
'
CM - = o O k (8)
) (
' T
CM - = k O k (9)
n
O ,..., , ) (
2 1
= = (10)
} ,..., , {
2 1 n
= (11)
<
=
>
=
0 1
0
0 1
i
i i
i
i
if
if
if

(12)
where o is a set of partially corrupted bipolar bits
generated by the decryption process, O is a threshold function
of hetero-correlation, represents multiplication result of the
correction matrix for the given distortion bit patterns, is set
of the recalled bits,
'
k represents result of exemplars and k is
a sequence of corrected bits.
Step 3: After performing error correction process, find out
the weighted distance between corrupted and corrected
exemplar as
(
(
=

=
n
i
i i
1
' '
k o u (13)
If 0 = u then distance becomes zero and engine decides
that the equilibrium point is reached, i.e., corrupted bits in
decryption process are safely recalled by hetero-correlators.
If s u , then assign corrected bits to
) ' ( ), ( ., . ,
'
k o k o o = = e i and perform step 2 until
distance of exemplar becomes zero.
If u > , then the engine confirms that adversary does
the correction process, therefore system has been terminated.
The is a vigilance parameter and it is calculated as
)) 2 , mod( ( n n =
i.e., )) 2 , mod( ( 0 n n s s and n
represents number of bits in an exemplar. The parameter
provides minimum energy for the bits correction between
genuine subject and partially corrupted cipher bits and also it
prevents local minima of the system. This parameter also gives
more complexity to the impostor to get back the original
messages.
Finally, recalled bipolar bits are converted to its equivalent
binary bits. These sequences of corrected bits represent the
original bits. The number of error bit recovery is based on
and E parameters. If 7-bit exemplar is used, then the
parameters 6 = and } 2 , 3 , 2 { = E provide a better result in
the error correction process.
V. EXPERIMENTAL RESULTS
The proposed approach has been implemented and
results were analysed. Efficacies of SIC and NRIC have been
evaluated. The NRIC systems time complexity was measured,
in that there were no recalling processes involved since the
encrypted bits were decrypted by the enrolled iris key. Hence
its enciphering and deciphering process depends on the time
complexity of iris-matching algorithm.
Next, the performance of the NRIC system was measured
by computing the time complexity of auto and hetero-
correlators recalling and encryption/decryption processes. In
the next experiment iris key energy complexities was
calculated in the case of cracking the messages by the
impostors. Finally, the guarantee issues of getting back original
bits were evaluated with respect to the energy variation of auto
and hetero-correlators. The detailed description of each
experiment is discussed in the following sections.
A. Speed performance
Time complexities of encryption and decryption process
have been evaluated for the SIC system. In that decryption
process required more time than encryption process, since the
decryption was performed after extracting and matching the iris
features at one time. The complexity of iris matching algorithm
was dependent on the size of the iris keys present in the system.
The complexity of searching iris keys iris key matching
system with linear search is O(N) and with binary search is
O(log N). The NRIC system required slightly more time than
the SIC approach because of its error correction engines require
more time to predict the original patterns from the partially
corrupted patterns. The search time of encryption and
decryption processes of SIC and NRIC are illustrated in Fig. 4.
Vol. 2, No. 11, 2011
6 | P a g e

Figure 4. Encryption and decryption time complexity of SIC and NRIC.
B. Recalling time
The recalling time of auto and hetero correlations
were dependent on size of the connection matrix in the error
correction process. The connection matrix was formed based
on the number of bits processed by the cipher text. In
accordance with the number of patterns and bits per exemplar,
the recalling time of auto and hetero correlators were evaluated
and shown in Fig. 5.

Figure 5. Auto and hetero correlators recalling time.
C. Performance issues
The guarantee issue of recalling process for correlators was
associated with two factors such as connection matrix of the
error correction engine and artifacts occurring on the iris
patterns. It provides nearly 97% of recalling entire pair of
trained patterns because of its local minimum of the energy
surface. However, in this paper, vigilance parameter was used
to put off local minimum attained by the system, i.e., energy
for the bits correction in between genuine subject and partially
corrupted cipher bits were computed to prevent the local
minima of the system. This parameter also produced more
complexity to the impostor to get back the original messages.
The factors of artifacts are fully concerned with three
possessions such as acquisition time users co-operation, non-
iris fractions occurring on iris and artifacts emerging in the core
area of iris. The guarantee issues of error correction process for
auto and hetero correlators are based on number of patterns and
bits per patterns used in the error correction process.
The guarantee performance of recalling process was
evaluated based on the Hamming distance between the
corrected bits and trained pairs. Multiple training was used to
recall several patterns. In this training, if pattern was not
recalled by the connection matrix by satisfying vigilance
parameter then train the patterns again by changing energy
constants, form a new connection matrix and performing
recalling process. This process was repeated until recalling
entire patterns by checking vigilance parameter. However,
trained patterns require sufficient number of bits to increase the
percentage of accuracy. Figure 6 shows the accuracy of
recalling patterns using auto and hetero correlators.

Figure 6 Accuracy of recalling patterns using auto and hetero correlators.
D. Impostor complexity
The probability of the presence of errors in the non-
repudiation process was assessed based on the number of bits
variation. These variations occur due to the environment,
illumination, occlusion of eyelids/eyelashes and other artifacts.
In this experiment, the number of bits corrupted in different
sessions was studied and verified in which situations brute
force search by an intruder can crack the iris crypto key. For
the experiment, different eye images were captured at different
sessions from the same subjects and their changes measured.
Figure 7 illustrates the error bit variation in different criterion.
The changes in bits may not be stable for all kind of capturing
because due to diverse changes the random alteration of bits
was assorted. The efficiency of the iris cryptosystem was
evaluated in accordance with key stability and strength. The
strength of the key can be evaluated based on entropy
principles. If message source alphabet was a2} {a1, A= and
the symbol probability P (a
1
) = 0.088 and P (a
2
) = 0.103 then
the entropy of the source symbol was 0.6495 bits/symbol. If an
intruder can tap the message, the probability of retrieving the
original message was ranged from 0.2 to 1 based on the error
Vol. 2, No. 11, 2011
7 | P a g e
bits of iris code. That is, if n bits were error then 2
n-26
times of
complication for brute force search was made to an intruder.
Thus the retrieving of the original messages has been made
complicated to the impostors. It provided a high key strength
for any cryptography system. This key cannot be stolen or
missed and gave more stability to the cryptosystem. These
types of bio keys can be produced every time the users want to
communicate secretly at non-secure channels. In addition,
experimental results show that this approach could easily be
adopted in the on-line cryptography systems as well.
E. Re-enrolments
Another design issue of integrating biometrics with
cryptography is the re-enrolments because biometric
cryptosystem is a reliable alternative for password protection
while releasing or direct usage of biometric key as a
cryptography key.

Figure 7. Error bit variation for the same subject in different criterion.
Hence encryption algorithm needs efficient solutions,
which are periodically updated biometric templates. Thus user
can register their patterns once in a month or other period of
time maintained in the system. Since some of the system
exploits biometric key for safeguarding mathematical
cryptographic keys or others may utilize as a part of the
biometric template. Nevertheless if biometric databases are
permanently stored in the local workstation for a period of
time, which is not secure, a system should employ the recently
enrolled iris keys for encryption process that increases the
system security and avoids eavesdropper attacks than the
lifelong biometric templates.
Thus the iris-based cryptosystem performs better accuracy
by using re-enrolments. In this paper, subjects iris patterns
were periodically enrolled once in a week in order to measure
the stability of the iris keys. However the keys variation
weighted distance was ranging from 0.0 to 0.19. This range
was fixed by statistical measures of iris recognition algorithm.
Thus these random variations were due to artifacts or other
non-iris sources. However the periodic amendment of genuine
subjects iris key produced more brute force search to the
impostors than the ordinary system.
VI. CONCLUSION
This research paper suggests a novel approach for iris based
cryptography system. The crypto keys have been generated
using iris patterns, which is stable throughout a persons
lifetime as well. Its inter-class variability for a person is very
large since it creates more complexity to crack or guess the
crypto keys. This approach has reduced a complicated
sequence required to generate keys as in the traditional
cryptography system. It can also generate more complex iris
keys with minimum amount of time complexity, which is aptly
suited for any real time cryptography system. This resolves the
key repudiation problem occurring in the traditional system.
The hetero-correlators can predict the number of bits corrupted
in the decryption process with the help of vigilance parameter.
The performance of the proposed approach is found to be
satisfactory.
In near-future, multi-modal cryptosystem will be
suggested to integrate biometric template to increase degree-of-
security in the non-secure data transmission.
REFERENCES
[1] Albert Bodo, Method For Producing a Digital Signature with Aid of a
Biometric Features, German patent DE 42 43 908 A1, 1994.
[2] Davida G.I., Frankel Y. and Matt B.J., On enabling secure applications
through off-line biometric identification, Proc. of IEEE Symposium
Privacy and Security, Oakland, California, USA,
pp. 148-157, 1998.
[3] Davida G.I., Frankel Y., Matt B.J. and Peralta R., On the relation of
error correction and cryptography to an offline biometric based
identification scheme, Proc. Workshop Coding and Cryptography
(WCC99), PARIS (France), pp. 129-138, 1999.
[4] Linnartz M.G. and Tuyls P., New Shielding Functions to Enhance
Privacy and Prevent Misuse of Biometric Templates, AVBPA 2003,
Guildford, UK, pp. 393-402, 2003.
[5] Clancy T., Kiyavash N. and Lin D.J., Secure Smartcard-Based
Fingerprint Authentication, Proc. of ACM SIGMM workshop on
Multimedia, Biometric Methods and Applications, New York, USA, pp.
45-52, 2003.
[6] Monrose F., Reiter M., Li Q. and Wetzel S., Cryptographic key
generation from voice, Proceedings IEEE Symposium on Security and
Privacy, Oakland, California, pp. 201-213, 2001.
[7] Yao-Jen Chang, Wende Zhang and Tsuhan Chen, Biometrics-Based
Cryptographic Key Generation, IEEE International Conference on
Multimedia and Expo, Taipei, Taiwan, (0-7803-8603-5/04), pp. 2203-
2206, 2004.
[8] Uludag U., Sharath Pankanti, Salil Prabhakar, Anil K. Jain, Biometric
Cryptosystems: Issues and Challenges, Proceedings of the IEEE, Vol.
92, No. 6, pp. 948-960, 2004.
[9] Daneil Schonberg and Darko Kirovski, Iris compression for
Cryptographically Secure Person Identification, Proceedings of IEEE
Data Compression Conference (DCC2004), Snowbird, UT, USA, pp.
459-468, 2004.
[10] Feng Hao, Ross Anderson and John Daugman, Combining
cryptography with biometrics effectively, Technical report of
University of Cambridge, No. 640, pp. 3-17, 2005.
[11] Bremananth R and Chitra A, An efficient biometric cryptosystem using
autocorrelators International Journal of Signal Processing 2;3, pp.158-
164, 2006.
[12] R.Bremananth and A.Chitra, A novel approach for high authentication
based on Iris keys, World Scientific and Engineering Academic Society
Transaction on Information science and applications, Issue 9, Vol. 2,
pp.1420-1429, 2005.
Vol. 2, No. 11, 2011
8 | P a g e
[13] R.Bremananth, A.Chitra, A new methodology for person identification
system, Sadhana, Indian academy of Sciences, Vol.31, Part 3, pp.259-
276, 2006.
[14] R.Bremananth, A.Chitra, Rotation Invariant Recognition of Iris,
Journal of Systems Science and Engineering, Vol.17, No.1, pp.69-78,
2008.
[15] Bart Kosko, Bidirectional Associative Memories, IEEE Transactions
on systems, Man, and Cybernetics, Vol. 18, No. 1, pp. 49-60, 1988.
[16] Yeou-Fag Wang, Jose B. Cruz and James H. Mulligan, Two coding
strategies for bidirectional associative memory, IEEE Transactions on
Neural networks, Vol. 1, No. 1, pp. 81-92,1990.
[17] Yeou-Fag Wang, Jose B. Cruz and James H. Mulligan, Guaranteed
recall of all training pair for bi-directional associative memory, IEEE
Transactions on Neural Networks, Vol. 2, No. 6, pp. 559-567,1991.
[18] P.Arul, A.Shanmugam, Generate a Key for AES using Biometric for
VOIP Network Security, Journal of Theoretical and appllied Information
Technology, Vol.5, No. 2, pp. 107-112, 2009.
(http://www.jatit.org/volumes/research-papers/Vol5No2/2Vol5No2.pdf)
AUTHORS PROFILE
Bremananth R received the B.Sc and M.Sc.
degrees in Computer Science from Madurai Kamaraj
and Bharathidsan University in 1991 and 1993,
respectively. He obtained M.Phil. degree in Computer
Science and Engineering from GCT, Bharathiar
University, in 2002. He received his Ph.D. degree in
2008 from Department of Computer Science and
Engineering, PSG College of Technology, Anna
University, Chennai, India. He has completed his Post-doctoral Research
Fellowship (PDF) from the School of Electrical and Electronic Engineering,
Information Engineering (Div.) at Nanyang Technological University,
Singapore, in 2011. He has 18+ years of experience in teaching, research and
software development. Currently, He is an Assistant Professor in the
Information Technology department, Sur University College, Sur, Oman,
affiliated to Bond University Australia. He received the M N Saha Memorial
award for the best application oriented paper in 2006 by Institute of Electronics
and Telecommunication Engineers (IETE). His fields of research are acoustic
holography, pattern recognition, computer vision, image processing,
biometrics, multimedia and soft computing. Dr. Bremananth is a member of
Indian society of technical education (ISTE), advanced computing society
(ACS), International Association of Computer Science and Information
Technology (IACIT) and IETE. He can be reached at
bremresearch@gmail.com.
Ahmad Sharieh had two bachelor degrees: one in
Mathematics and one in Computer Sciences. He had
master degree in Computer Science and High Diploma in
Teaching in Higher Education. He had PhD in Computer
and Information Sciences from Florida State University
1991. Sharieh worked as Assistant Professor in Fort
Valley College / USA and The University of Jordan (UJ)
/ Jordan. He worked as Associate Professor with Amman
Arab University for Graduate Studies (AAUGS) / Jordan and The University of
Jordan. He worked as Dean of King Adbullah School for Information
Techology/Jordan. Currently, he is a professor of Computer Sciences and Dean
at Sur University College (SUC), Oman. He published articles in journals (27),
in conferences (22), and authored and prepared books (14). He gained grant for
eight research projects from UJ and Europe. He developed several software
systems such as: Teaching Sign Language, e-learning Modeling and
Simulation, and Online (Automated) Exams. He is on the editorial board of
several journals and conferences and a referee of several others. His research
areas are Distributing Systems, Expert Systems, E-Government, E-Learning,
Parallel Processing, Pattern Recognition, Software Engineering, Wire/Wireless
Communication, Modeling and Simulation.

Vol. 2, No. 11, 2011
9 | P a g e
On the transmission capacity of quantum networks

Sandra Knig
Alpen-Adria Universitt Klagenfurt
9020 Klagenfurt, Austria
Stefan Rass
System Security Research Group
Alpen-Adria Universitt Klagenfurt
9020 Klagenfurt, Austria

AbstractWe provide various results about the transmission
capacity of quantum networks. Our primary focus is on
algorithmic methods to efficiently compute upper-bounds to the
traffic that the network can handle at most, and to compute
lower-bounds on the likelihood that a customer has to wait for
service due to network congestion. This establishes analogous
assertions as derived from Erlang B or Erlang C models for
standard telecommunications. Our proposed methods, while
specifically designed for quantum networks, do neither hinge on
a particular quantum key distribution technology nor on any
particular routing scheme. We demonstrate the feasibility of our
approach using a worked example. Moreover, we explicitly
consider two different architectures for quantum key
management, one of which employs individual key-buffers for
each relay connection, the other using a shared key-buffer for
every transmission. We devise specific methods for analyzing the
network performance depending on the chosen key-buffer
architecture, and our experiments led to quite different results
for the two variants.
Keywords-Quantum network; Quantum cryptography; network
transmission capacity; queuing network; system security.
I. INTRODUCTION
It took about two decades ever since quantum key
distribution (QKD) has been proposed by [1] (the famous
BB84 protocol) until the first experimental implementations of
a quantum network were presented by the DARPA [2] and the
European Union [3]. While the theory behind secure key-
delivery between Alice and Bob is well-understood (see e.g. [4]
for a proof regarding the security of BB84), the theory of
network design and performance analysis has apparently seen
rather limited attention over the last years. The works of [5]
and [6] both considered the design of a network from the
topological point of view, and in terms of optimal security and
performance. In this work, we go the other way, asking for the
best performance that we can get from a given quantum
network infrastructure. In particular, we provide algorithmic
means to answer two questions:
1. What is the maximal transmission capacity achievable
in the network (using any classical routing scheme)?
2. What is the likelihood of local congestion that would
temporarily disconnect the (logical) channel between
any two peers in the network?
The second question can be rephrased into asking how
likely a customer is to wait when asking the network for a
secure delivery of payload from one point to another.
Our results are hence related to the field of communication
theory, channel capacity and network coding. Particularly the
latter has led to valuable insights (cf. e.g. [7], [8], [9])
regarding the rate at which information can be send through the
network. Contrary to these (and many other related)
approaches, we do not employ classical information theory to
quantify the capacity, but rather work with the directly known
performances of each link in the network. Similar to our work,
network outage probabilities are as well discussed in [10], [11],
[12] and [13], where most research effort, as it seems, has been
put on wireless networks. So far, the problem appears hardly
considered in (hard-wired) networks or quantum networks. In
the quantum computing domain, the work of [14], [15], [16]
and [17] is closely related. Contrary to ours, however, it
strongly relies on quantum techniques and is less focused on
algorithmic methods to analyze a given infrastructure. The
work of [16] is particularly interesting as it employs
percolation theory (which is rarely used in the related
literature). The problem is studied elsewhere in [18], which
comes up with proposals on how to enhance the existing
capabilities once they are known. Here, we work out the limits
similarly to the Erlang B and Erlang C models, so as to be able
to improve them based on this related research. In the wireless
domain, the interesting work of [19] deals with spread-
spectrum techniques and uses Poissonian processes for
determination of the network capacity, but is specific for this
particular encoding technique. We explicitly avoid such
restrictions here, but adopt some assumptions on the quantum
key-management models (following the proposals of [20]; cf.
also [21] for another discussion related to quantum key- and
network-management). Finally, we mention the work of [22],
who attempts to solve the problem of end-to-end quantum
communication using a three-party protocol. This architecture
is essentially different from what has been implemented
nowadays, and thus subject of future considerations.
Among the sales arguments for quantum key distribution is
its capability of running over existing fibre-optic lines. This
claim has been substantiated in the demonstration of the
SECOQC- and DARPA-Networks [2], [3]. Hence it is fair to
assume that the topology of the network is fixed and that the
individual key-generation rates on each link are known (cf. [3]
for examples). Moreover, following the so-far proposed
architectures of relay devices in a quantum network (see [2],
[3],and [20]), it is reasonable to consider a quantum network as
a system of interconnected buffers, where the secret message is
repeatedly decrypted and re-encrypted before forwarding it to
the next hop. The key-buffers in each relay node are constantly
refilled by QKD protocols running in the background 24/7 and
Vol. 2, No. 11, 2011
10 | P a g e

Figure 2 Network with link performances

Figure 1 Trusted relay with re-encryption in each hop
endlessly generating key-material for later usage. This
transmission regime, in its simplest form, is known as trusted
relay, and is widely used in nowadays demonstration networks
(cf. [2] and [3]). A transmission of a message from Alice to
Bob along a sequence of nodes that share keys
is
displayed in Figure 1.
Organization of this work: in Section II, we describe the
graph- and queuing model (Section II.A) used to analyze a
quantum network. In particular, we will use maximal flows
(briefly introduced in Section II.B) to compute bounds on the
payload that the network can handle. Section III is divided into
two main parts, giving algorithms for computing end-to-end
transmission capacity (Section III.A) and waiting probability if
a chosen path through the network is blocked due to congestion
(Section III.B). In both cases, we show how to use standard
maximum flow and shortest path algorithms to compute the
desired quantities easily and efficiently from the known link
capacities in the underlying graph model. The process is
illustrated by an example (Section III.C), before conclusions
follow in Section IV.
II. MATHEMATICAL MODEL
We will not burden ourselves with the details of any
particular quantum key distribution (QKD) facility, but restrict
our attention to the following model of a quantum network
(QNet): let a QNet be modeled as an undirected graph
with adjacent nodes sharing secret keys thanks to QKD.
That is, on any line (with ) maintain key-buffers
on either side to store QKD key material for subsequent
transmissions. These key-buffers are nothing else than queues,
in which key-bits are inserted on a deterministic basis (we
assume the QKD-devices to generate key-bits at constant rate).
key-bits are used up on a random basis, depending on incoming
payload for secret transmission. Based on this view, we can
cast the QNet into a standard queuing network.
A. Quantum Networks as Queuing Networks
An open queuing network is a system in which a customer
enters the network at some node, and moves onwards through
the links, where he occasionally has to wait (queue) until he is
served at the next node (i.e. he can enter the next node). Central
questions in queuing theory regard the average time to wait
until the customer reaches his destination point, or the average
number of customers lining up in any given queue (link) in the
network. For a quantum network, we can equally well set up
such a model, based on the following correspondence:
1. incoming customers equal newly generated key-bits
2. leaving customers equal the (one-time) use of key-bits
for Vernam-encryption of messages
3. a queue equals a key-buffer (storing bits for subsequent
usage, or equivalently, hosting customers for
subsequent service)
Observe that the generation of key-bits is deterministic,
while the arrival of messages is non-deterministic. If we
consider the QNet as a backbone network, then it is reasonable
to consider the event of an incoming message bit as a Poisson-
distributed random variable. In Kendall-notation, the link
in a QNet therefore is nothing else than a -queue,
disregarding physical size limits of the key-buffers for now.
The graph modeling the QNet thus constitutes a network of
queues, and we are interested in its stationary distribution (so it
exists).
Remark: an alternative view yielding equivalent results is
by associating incoming payload bits with customers, who get
served (encrypted) based on how much key-material is
available in the buffer. In this case, we would have to think of
the message-queue as the physical incarnation of the queue
model under consideration. For simplicity, and because real-
life implementations work with a key-buffer too, we shall
consider the first of these two views, keeping the second view
in mind whenever needed.
Sufficient conditions for the existence of stationary
distributions in queuing networks are well-known, such as
Jackson's theorem for Jackson-networks or one of its
generalizations, such as the BCMP-theorem. Openness of the
network is assured, since the graph is a mere transportation
medium and messages necessarily leaving the network at some
stage. Still, we cannot make any generally valid assertions on
the routing algorithms implemented within the system. In lack
of such information, we will try to find upper bounds to the
transmission capacity by invoking maximum flow theory. This
has the additional advantage of our results applying to
conventional routing as well as network coding approaches for
transmission.
Vol. 2, No. 11, 2011
11 | P a g e

Figure 3 Shared vs. individual key-buffers

Figure 4 Forwarding with link-specific keys
B. Flows
To illustrate the approach, consider the example network
topology shown in Figure 2. Assume that after start-up, all key-
buffers are empty and the QKD-protocol on link
start
producing key-material at constant rate
per time unit. It

follows that after one unit of time, the maximal transmission
capacity of the network is determined by the maximum -
flow, constrained by the existing key-material on each link. As
all links regenerate key-bits at constant rate, the minimum cut
will not change over time. Let be the minimum edge
cut, then the maximum flow has capacity
)
after the first unit of time. After units, we have
the capacity
The
consumption of key-bits happens upon arrival of payload to be
transmitted secretly from the sender to the receiver . Hence,
we can consider the entire network as one large queue, whose
internal servicing is done by routing, network coding, or
otherwise. From outside, we have
as the deterministic
rate at which key-bits arrive for later consumption, and we are
back at the -queue. Considering multiple access-points
to the network is trivial by switching to a multi-source-multi-
sink flow. Unlike standard queueing disciplines, optimality in
our context means the incoming amount of key-material
outweighing the arrival of messages, i.e. an "unstable" queue
whose expected length is infinite
C. Key-Buffer Architectures
It is easy to set up the devices so as to realize a single-path
transmission as illustrated in Figure 1. However, it would not
be reasonable to assume nodes to have only two ports, so the
internal management of quantum keys is slightly more
involved. Going back to Figure 1, we can instantly fix the
problem of the message popping up in plaintext within each
relay node by simply XOR-combining both, the incoming- and
outgoing key into a single relay-key [20]. Figure 4 illustrates
the idea: for the relay from node A to node B over node R, the
latter would XOR-combine (-operation)
and
into
.
Consequently, we would only store
in an individual buffer
for this link (see Figure 3; right). If the relay is trusted, then it
may alternatively decrypt the incoming message and re-encrypt
it before passing it onwards (as shown in Figure 1).
Consequently, we would have to maintain shared buffers for
each I/O-port, as displayed in Figure 3 (left) .

III. RESULTS
We are now ready to present our main results.
A. End-to-End transmission capacity
Given the (constant) rate of key-bits generated on link
, we can ask for the maximal average rate of arriving
messages that we are able to encrypt. Or stated differently: if
we know the service rate, what is the highest rate of incoming
customers that we can handle? Obviously, a necessary
condition is the rate of arriving customers not exceeding the
service rate. This is almost sufficient, as the following result
shows:
Proposition 1. Let a -queue be given and denote the
average arriving rate by . For a given constant service
rate , any arrival rate leads to a stable system.
So although the case may be fine for a system with
deterministic arrival, it is not appropriate for random systems.
Proof: Denote by the number of customers in the system
and by
the probability of being in state . In a

stable situation, the total rate out of this state equals the sum
of the incoming rates from states and , which
yields for every ,

.
This recursive formula can be rewritten as

and since probabilities sum up to one we get

This is simply a geometric series, which converges if and
only if , as stated.

The above proof also shows that the probabilities
can be
calculated via
for any (by using

where
.
Putting this to practice within a quantum network is
straightforward in two steps:
1. Upon given key-generation rates on each link, use
Vol. 2, No. 11, 2011
12 | P a g e
these rates as edge-weights in an undirected graph and
determine a maximal flow from the source node to the
sink node. Call the value of this flow .
2. By Proposition 1, a payload of up to bits per
time unit can be handled by the quantum network in a
perfectly secure manner (assuming trusted relay).
B. Waiting probability
Here, we need to distinguish two architectures in our
theoretical considerations: let a node be given, whose
neighbors are

eg
, where denotes s
degree. Either each route passing through is associated with
its individual key-buffer (perhaps via logically partitioning the
overall key-material somehow; cf. the right side of Figure 3), or
all incoming and outgoing flow draws from the same key-
buffer (Figure 3; left), in which case a very busy line can affect
the capacities of other routes through . However, short-term
traffic peaks are easier to handle with this architecture. We
consider both variants separately.
Individual key-buffers
To estimate the probability that one has to wait to get the
key-bits needed for encryption anywhere along its way from
the sender to the receiver, we first focus on the waiting
probability in one particular node . Observe that since the
key-buffers are not shared, distinct links from a node to any
of its neighbors act independently. So we can restrict our
considerations to any (arbitrary) key-buffer within . Notice,
however, that a link from to its neighbor has to be treated
differently than the link from to , since we are concerned
with forwarding packets.
Let be any node, and denote its neighbors by
. Pick any key-buffer within that refers to the
connection , where . Let the incoming traffic
per time unit on the route from to be
distributed, and assume the QKD protocol between and to
reproduce an amount of bits per time unit. Finally, assume
the key-buffer to be full at the beginning. Let
be a
sequence of i.i.d. random variables
where

is the traffic at time unit . The corresponding filling level of
the key-buffer at time can never exceed the capacity and is
thus given by

} (1)
assuming that we start off with the full key-buffer and re-fill it
at rate after the first time-unit (i.e. we do no refill within the
first time-unit because the buffer is full already).
We are interested in the probability for the link being
blocked, i.e. the likelihood of an empty key-buffer at time unit
. Hence, we ask for
. From
we deduce {
} if
and only if
. Hence,
{
}
{
}
If the traffic load over different time-units is independent,
then
so that the above probability

boils down to a mere evaluation of the Poisson distribution
function |
and comes to
|
It is legitimate to ask what happens if the refilling of the
key-buffer happens on a random basis as well. Call
the
amount of fresh key-material in time-unit . We can easily
replace the term in (1) by
so as to take this
randomness into account, but the distribution of
is
no longer Poissonian (mostly because the difference is not
bounded from below). A straightforward way out of this
dilemma is considering Gaussian approximations to the
Poissonian densities, which takes us back to the wonderful
world of distributions closed under convolution. In other
words, if we approximate
by
the above derivation and result becomes

obvious. We leave this track aside here and go back to the
deterministic refilling, giving us the following result:
Proposition 2. Let have neighbors , and
consider the key-buffer shared with an arbitrary but fixed
neighbor of . Denote by
the size of the

key-buffer associated with the link and assume that
new key-bits are generated at a constant rate . The
number of incoming message bits from to is assumed
to be -distributed. If , then the
probability of waiting within during a transmission is
oraring rom to gets elaye

where

Proof: The argument is merely a matter of noticing that it
takes a period of
time units to entirely empty the key-

buffer, if bits are used up with bits growing back per time
unit (the time for encryption is considered negligible). Hence,
the average expenditure is .
Vol. 2, No. 11, 2011
13 | P a g e

Figure 5 Transformation of a node with individual key-buffers
The alert reader might utter concerns about the stochastic
independence of incoming traffic over different time-units.
There are (at least) two ways to justify this assumption:
Considering the transmission as a process resting on
various routing protocols, we can reasonably assume
the network's routing regime to rearrange, encode and
decode, and to partition the messages in a way that
stochastic correlations between packets are negligible.
Notice that this in no way rules out the possibility of
linking packets to each other via sequence numbers or
matching delivery addresses. However, this "weaker"
type of correlation does not necessarily imply
dependencies among the payloads of different packets,
e.g. when a long sequence of independent
cryptographic key-material is transmitted.
In case that the network is merely used for continuous
shared key establishment (in fact, this is the way in
which a quantum link is generally supposed to be used
[3]), we can safely assume incoming traffic packets as
stochastically independent, for otherwise we would
have interdependence among key-bits. This is
undesired for cryptographic keys, particularly for
quantum keys (as it reduces the key's entropy).
While Proposition 2 refers to only a single node, it is
more interesting to find out how likely a blockage
along a path from any node to any other node is. In the
model of individual key-buffers, this problem boils
down to identifying a path whose blockage probability
is minimal. We can simply invoke any shortest-path
algorithm for that matter, if we assume blockages to
happen independently of each other. Consider a node
having neighbors .
Observe that Proposition 2 is concerned with the likelihood
of blockage when forwarding a message from onwards.
Hence, we need to cast the undirected network model graph
into a directed graph by converting an undirected edge into two
directed edges (with opposite directions).
Each link for maintains its own key-buffer
with blocking probability

as given by Proposition 2. We
transform the undirected graph into a weighted
directed graph such that
1. Each link is carried over into two links
with cost .
2. Each node having a number || of
neighbors is expanded into a complete graph with
nodes
, each of is connected by two directed

edges in either direction. Each edge
is added to
with the cost

according to Proposition 2. The set of edges joining
to its neighbors in , i.e. the set { }

is carried over to as
with
weights all zero.
Figure 5 illustrates this transformation.
On , we can run any shortest-path algorithm, to
determine the minimum likelihood of blockage using
single-path routing. For any given sender and receiver ,
let their most reliable interconnecting path have "weight"
in . Then, regardless of the routing, we have
message ill e loke
because is the weight of the shortest path in , i.e. the
most reliable path in . No matter what the routing actually
does, it cannot do better than choosing the best path
possible, hence the value
at least one noe is loke
is a lower-bound to the actual likelihood.
Shared key-buffers
In the case of shared key-buffers we assume the
incoming flows from different nodes to be independent.
Then the distribution of the total flow trough at
time follows a Poisson distribution with parameter

, where
denotes the incoming traffic flow

from node to its neighbor . Similarly, all neighboring
links contribute
bits of fresh key-material to

the common key-buffer, giving a total refreshing rate of

. With we can invoke proposition
Proposition 2 again to calculate the probability of a node
being blocked in this case.
A little more care is to be taken when asking for the
chance of blocking somewhere across the network as a
whole. In this case, we use a transformation that is normally
used to calculate maximal flows with vertex capacities. The
transformation from the undirected graph (see
Figure 6a for an example) to the directed weighted graph
is now specific for a sender and receiver ,
and proceeds as follows (cf. [23]):
1. Each node including and is replaced by two
nodes
in

out
, and a directed edge from
in
to
out

is placed to . This link
in

out
gets assigned the
cost , where is the blocking probability
calculated as described above.
2. Each undirected edge is replaced by two
directed edges
out

in
and
out

in
. See Figure
6b and Figure 6c for an illustration.
3. The nodes
in
and
out
are deleted, as well as all edges
going into
in
and out of
out
.
4. Those nodes who remain to be assigned a cost receive
zero cost. The resulting graph is shown in Figure 6d.
Vol. 2, No. 11, 2011
14 | P a g e

Figure 7 Example network (link performances shown in kbit/sec)
Once having found the shortest path in between and ,
we can draw exactly the same conclusion as above: if is the
weight of this path in , then the chance of this path being
blocked for any path-based routing-scheme is lower-bounded
by
.
C. Example
To get a more intuitive understanding of the above results,
we give a simple example. Consider a modified version of the
graph from Figure 2, with six nodes but with neither an edge
between vertices 2 and 3 nor between 2 and 4. We call node 0
the sender and the receiver shall be node 5 (cf. Figure 7). We
let be 5000 kbits and choose the rates (at which new key-
material is produced) randomly between 280 and 320 kbits and
let the rate of the incoming message bits be (see
Table 1(a)), where is a positive constant ( in this
specific example). Under this setting, we get the average
probabilities
shown in Table 1(a) for the incident of

waiting between and where the average was taken over
calculations.
Using the transformation described above we get that the
probability of getting stuck is lower bounded by 0.9137 in the
case of individual key-buffers and by 0.9929 in the case of
shared key-buffers; again averaged over trials. This
means that the individual link performances are indeed sharp
bounds to the true bandwidth, as even slightly overshooting (by
kbit/sec in out example) makes congestions highly
likely. If we just look at a single evaluation of the two different
methods mentioned above, we also see that the paths yielding
the minimal value may differ: while working with the design of
individual buffers the algorithm takes route in
the original graph, it does prefer under the
shared buffer design.
An illustration and interpretation of Proposition 1 is the
following: with link performance values as given in Figure 7, a
maximal flow is found at 607.54 kbit/sec. So this is the
maximal traffic load that the network can handle.
Better performance is obtained when we double the size of
the key-buffer in each link. Under the same set up as before,
but with Mbit of key-buffer and the -values as listed
in Table 1(b), we get the blocking probabilities shown in the
right column of Table 1(b). The blocking probability for an
end-to-end communication in this case is when
individual key-buffers are used, and when a shared
buffer is employed. Finally, Proposition 1 tells that the overall
end-to-end traffic from 0 to 5 is bounded above by 608.52
kbit/sec (which is the maximal flow under the respective
capacities
for each link).

It is important to observe that Proposition 2 explicitly is
concerned with situations in which the traffic load exceeds the
key-(re-)generation rate on the links. The converse case in
which there is a positive surplus of key-material produced on
each link is obviously not interesting in terms of congestion
likelihoods (as the key-buffers cannot run empty in that case).

Figure 6 Transformation for shared key-buffer
Vol. 2, No. 11, 2011
15 | P a g e
TABLE 1 BLOCKING PROBABILITIES EXAMPLES
Edge Traffic [kbit/s] Blocking prob.

305.23
0.7062

302.32
0.7058

304.70
0.7063

305.35
0.7067

305.77
0.7066

304.28
0.7064

305.65
0.7068

304.22
0.7064
(a) Key-buffer size Mbit
Edge Traffic [kbit/s] Blocking prob.

306.06 0.6497

304.53 0.6493

304.29 0.6492

305.29 0.6495

305.89 0.6499

304.34 0.6494

305.65 0.6497

304.23 0.6493
(b) Key-buffer size Mbit
IV. CONCLUSIONS
Given a quantum network, we have shown how to
efficiently compute bounds to the transmission capacity and the
likelihood of blocked paths due to local congestions.
A. Future Work
We focused on two specific architectures for key-buffers.
Our approach and results are extensible towards more general
arhitetures (as e onsiere only to extreme ases here)
for the key-buffers as well as for the relay-regime as such (cf.
[22], who propose a novel three-party quantum communication
approach). It is well known that classical routing regimes face
difficulties when trying to attain the upper bounds to the
transmission capacity as implied by the max-flow approach
(network coding is one way to resolve this dilemma).
Consequently, our bounds are not necessarily tight. A closer
investigation of this is subject of future work. Finally, since
quantum networks have hardly reached a level of maturity
beyond prototypes or lab demonstrators, reports on
comparisons of our results to other competing approaches are
part of future research.
B. Summary
Our analysis is entirely based on the physical topology of
the network and the known key-generation rates on each link.
In this work, we focused on single-path (classical) routing
schemes, leaving analogous research in the field of multipath
routing and network coding for future work. Our results are
easy to implement with off-the-shelf algorithms, hence the
proposed analysis technique is efficient in terms of
computational, modeling and implementation efforts.
Despite quantum networks not having evolved beyond
demonstrator prototypes yet, the possibility of setting up a
high-security transmission network over existing fibre-optic
lines is quite interesting. Our research here is meant as a
starting point towards the construction of such infrastructures
in an effective and appealing manner for the potential
customer. Quality of service and service level agreements in
quantum networks, unfortunately, have by now not seen the
necessary attention to really bring the QKD technology to the
market. Although ingenious solutions and brilliant theoretical
achievements have been made, the "last mile" between lab
implementation and large-scale practical business
implementation needs more attention.
REFERENCES
[1] C. Bennett and G. Brassard, "Public key distribution and coin tossing,"
in Proceedings of IEEE International Conference on Computers Systems
and Signal Processing , Bangalore, India, 1984.
[2] C. Elliott, "The DARPA Quantum Network," arXiv:quant-
ph/0412029v1, 2004.
[3] M. Peev, C. Pacher, R. Alleaume, C. Barreiro, J. Bouda, W. Boxleitner,
T. Debuisschert, E. Diamanti, M. Dianati, J. F. Dynes, S. Fasel, S.
Fossier, M. Furst, J. D. Gautier, O. Gay, N. Gisin, P. Grangier, A.
Happe, Y. Hasani, M. Hentschel, H. Hubel, G. Humer, T. Lnger, M.
Legre, R. Lieger, J. Lodewyck, T. Lornser, N. Ltkenhaus, A. Marhold,
T. Matyus, O. Maurhart, L. Monat, S. Nauerth, J. B. Page, A. Poppe, E.
Querasser, G. Ribordy, S. Robyr, L. Salvail, A. W. Sharpe, A. J. Shields,
D. Stucki, M. Suda, C. Tamas, T. Themel, R. T. Thew, Y. Thoma, A.
Treiber, P. Trinkler, R. Tualle-Brouri, F. Vannel, N. Walenta, H. Weier,
H. Weinfurter, I. Wimberger, Z. L. Yuan, H. Zbinden and A. Zeilinger,
"The SECOQC quantum key distribution network in Vienna," New
Journal of Physics, vol. 11, no. 7, p. 075001, 2009.
[4] P. Shor and J. Preskill, "Simple Proof of Security of the BB84 Quantum
Key Distribution Protocol," Phys. Rev. Lett., vol. 85, pp. 441-444, 2000.
[5] R. Alleaume, F. Roueff, E. Diamanti and N. Ltkenhaus, "Topological
optimization of quantum key distribution networks," New Journal of
Physics, vol. 11, p. 075002, 2009.
[6] S. Rass, A. Wiegele and P. Schartner, "Building a Quantum Network:
How to Optimize Security and Expenses," Springer Journal of Network
and Systems Management, vol. 18, no. 3, pp. 283-299, 2010.
[7] S. Bhadra and S. Shakkottai, "Looking at Large Networks: Coding vs.
Queueing," in Proceedings of the 25th IEEE International Conference on
Computer Communications, Barcelona, Spain, 2006.
[8] S.-Y. R. Li, R. W. Yeung and N. Cai, "Linear Network Coding," IEEE
Transactions on Information Theory, vol. 49, no. 2, pp. 371-381, 2003.
[9] R. Ahlswede, N. Cai, S.-Y. Li and R. Yeung, "Network information
flow," IEEE Transactions on Information Theory, vol. 46, no. 4, pp.
1204-1216, 2000.
[10] D. H. Woldegebreal, S. Valentin and H. Karl, "Outage probability
analysis of cooperative transmission protocols without and with network
coding: inter-user channels based comparison," in Proceedings of the
10th ACM Symposium on Modeling, analysis, and simulation of
wireless and mobile systems, New York, USA, 2007.
[11] S. Weber, J. G. Andrews and N. Jindal, "An overview of the
transmission capacity of wireless networks," IEEE Transactions on
Communications, vol. 58, no. 12, pp. 3593-3604, 2008.
[12] M. Kaynia, F. Fabbri, R. Verdone and G. ien, "Analytical study of the
outage probability of ALOHA and CSMA in bounded ad hoc networks,"
in Proc. European Wireless (EW), 2010.
[13] R. Vaze and R. W. Heath, "Transmission Capacity of Ad-hoc Networks
with Multiple Antennas using Transmit Stream Adaptation and
Interference Cancelation," http://arxiv.org/abs/0912.2630, 2009.
[14] F. Caruso, S. F. Huelga and M. B. Plenio, "Noise-Enhanced Classical
and Quantum Capacities in Communication Networks," Phys. Rev. Lett.,
vol. 105, p. 190501, 2010.
[15] S. Watanabe, Remarks on Private and Quantum Capacities of More
Capable and Less Noisy Quantum Channels, arXiv:1110.5746v1 [quant-
ph], 2011.
[16] C. Le Quoc, P. Bellot and A. Demaille, "On the security of quantum
networks: a proposal framework and its capacity," in Proceedings of the
Vol. 2, No. 11, 2011
16 | P a g e
International Conference on New Technologies, Mobility and Security,
2007.
[17] G. Smith, J. A. Smolin and J. Yard, "Quantum Communication with
Gaussian channels of zero quantum capacity," Nature Photonics, vol. 5,
pp. 624-627, 2011.
[18] G. Zhang, S. Zhou, D. Wang, G. Yan and G. Zhang, "Enhancing
network transmission capacity by efficiently allocating node capability,"
Physical A: Statistical Mechanics and its Applications, vol. 390, no. 2,
pp. 387-391, 2009.
[19] S. Weber, X. Yang, G. d. Veciana and J. Andrews, "Transmission
capacity of CDMA ad-hoc networks," in IEEE Eighth International
Symposium on Spread Spectrum Techniques and Applications, 2004.
[20] P. Schartner and S. Rass, "How to overcome the Trusted Node Model in
Quantum Cryptography," in Proceedings of the 12th IEEE International
Conference on Computational Science and Engineering, 2009.
[21] A. Mink, L. Ma, T. Nakassis, H. Xu, O. Slattery, B. Hershman and X.
Tang, "A Quantum Network Manager That Supports A One-Time Pad
Stream," in Proceedings of the second International Conference on
Quantum, Nano and Micro Technologies, 2008.
[22] C. Le Quoc and P. Bellot, "A New Proposal for QKD Relaying Models,"
in Proceedings of 17th International Conference on Computer
Communications and Networks, 2008.
[23] A. Abbas, "A Hybrid Protocol for Identification of a Maximal Set of
Node Disjoint Paths," International Arab Journal Of Information
Technology (IAJIT), vol. 6, no. 4, pp. 344-358, 2009.
AUTHORS PROFILE
Sandra Knig received her Bachelor and Master degree in Mathematics at the
ETH Zurich, with a focus on statistics (subspecialty in regression
analysis and time series analysis). Her research interests cover
applications of statistics in electrical engineering as well as
communication and information theory .

Stefan Rass graduated with a double master degree in mathematics (with a
focus on statistics)and computer science from the Klagenfurt University
in 2005, and gained a PhD degree in mathematics in 2009. His research
interests include general system security, in particular applications of
quantum cryptography and information-theoretic security.

Vol. 2, No. 11, 2011
17 | P a g e
Confidential Deterministic Quantum Communication
Using Three Quantum States

Piotr ZAWADZKI
Institute of Electronics
Silesian University of Technology, POLAND

AbstractA secure quantum deterministic communication
protocol is described. The protocol is based on transmission of
quantum states from unbiased bases and exploits no
entanglement. It is composed form two main components: a
quantum quasi secure quantum communication supported by a
suitable classical message preprocessing layer. Contrary to many
others propositions, it does not require large quantum registers.
A security level comparable to classic block ciphers is achieved
by a specially designed, purely classic, message pre- and post-
processing. However, unlike to the classic communication, no key
agreement is required. The protocol is also designed in such a
way, that noise in the quantum channel works in advantage to
legitimate users improving the security level of the
communication.
Keywords- quantum cryptography; quantum secure direct
communication; privacy amplification.
I. INTRODUCTION
The interest in quantum communication is motivated by the
promise of provable security based on the laws of physics. The
most mature protocols use quantum channels for secure
quantum key distribution (QKD) which is further used by
legitimate parties to protect communication over classic
channels [1]. The content of the key resulting from QKD
execution is not determined by either of users but is random
and settled by the protocol completion itself. An alternative
paradigm of a quantum secure direct communication (QSDC)
has been investigated in the last decade [2,3]. QSDC protocols
offer confidential transmission of deterministic classic
information over a quantum channel without a prior key
agreement. Most QSDC protocols are completely robust [4-7].
It means, that an eavesdropper cannot intercept any information
without introducing errors in the transmission. Unfortunately,
the absence of privacy amplification step in QSDC protocols
causes that complete robustness guarantees only quasi security
in perfect quantum channels there exists finite, nonzero
probability that some information is intercepted without
detection. Situation is even worse in noisy environments when
legitimate users tolerate some level of transmission errors. If
that level is to high compared to the quality of the channel then
an eavesdropper can peek some fraction of signal particles
hiding himself behind acceptable QBER threshold. The
possibility to intercept some part of the message without being
detected renders protocol insecurity. This difficulty has been
resolved by processing qubits in blocks [8-12] and/or using
quantum privacy amplification [13,14]. However, such an
approach requires large quantum registers which are not
realizable at present with photonic techniques.
In this paper QSDC security is improved by message
classic processing. The quantum protocol based on single
photon transmission and not exploiting quantum entanglement
is supplemented by pre- and post-processing procedures. The
quantum part requires only one qubit register and such
photonic quantum memory operating in high temperature has
been already realized experimentally [15]. The preprocessing
part is an adaptation of the transform proposed in [16] to the
specific requirements of quantum communication. In the
resulting protocol, contrary to many others QSDC protocols,
the noise in quantum channel works in advantage for the
legitimate users improving the security of communication.
II. PROTOCOL DESCRIPTION
Alice, the sender of information, is able to generate three
quantum states , and . Bob,
the recipient of the message, is equipped with one qubit
quantum register and is able to perform quantum measurements
in Z and X bases. Users are connected with quantum and
classic communication channels. Information in classic channel
is not confidential and may be freely eavesdropped. However,
it is assumed that this information cannot be modified by
distrusted parties. On the contrary, the quantum channel may
be tampered with no limitations any data manipulation
allowed by the laws of physics is permitted. Let
and {
} be the data block of bits which Alice is going to

send.
A. Preprocessing
1. Alice generates a random sequence of bits {
}.
This sequence is further called a preprocessing key.
2. Alice encrypts some publicly known text T with the
classic cipher of a well established reputation that
produces a ciphertext {
of size
using a preprocessing key (
denotes encryption
operation),
3. The preprocessed sequence (
) which will
be sent is composed of two parts. The first part is
formed by bitwise xoring the ciphertext from the
previous step with the message bits:

, (1)
Vol. 2, No. 11, 2011
18 | P a g e
and the second part of sequence S is calculated as

. (2)
The encoding operation is invertible only when all bits of the
sequence S are received without errors, thus some error
correction code is used to protect against noise .
Sequence B is sent to Bob via the quantum channel.
B. Communication
Qubits are processed in one-by-one manner. Alice
randomly switches between control and message mode and
uses classic channel to notify Bob, that the last qubit of the
given data block was sent so he should proceed with the post-
processing step. The sequence {
} is sent first.
1) Control mode
1. Alice randomly prepares one of the states ,
or which is sent to Bob.
2. Bob stores the received state in quantum register and
notifies Alice.
3. Alice informs Bob that this qubit should be processed
in control mode and informs Bob about the state
preparation basis (Z or X).
4. Bob measures a quantum register in the basis
specified by Alice.
5. If the selected basis was X, Bob knows that the
measured state should be . The appearance of
denotes a transmission error. Alice is notified
about a failure.
6. If the selected basis was Z, Bob informs Alice about a
measurement result and Alice compares that result
with the value used in the state encoding. Bob is
informed about comparison correctness.
If an error rate averaged over sufficiently large number of
control qubits exceeds the correction capabilities of the ECC
code then transmission is aborted before entire message is sent.
2) Message mode
1. Alice encodes bit sequence {
} as states
and and sends them to Bob.
2. Bob stores the received state in quantum register and
notifies Alice.
3. Alice informs Bob that this qubit should be processed
in message mode.
4. Bob measures quantum register in Z basis, stores the
measurement result as
in classic memory and

notifies Alice that he is ready for the reception of the
next qubit.
C. Postprocessing
ECC data is used to correct errors on the received sequence
,
1. The preprocessing key is recovered as
.
2. The ciphertext sequence is again calculated as
and the message decoded as
.
If any of the bits in the sequence
is incorrect then the key

is also incorrect and the sequence
is completely different
from (this behavior is guaranteed by the properties of the
classic cipher). It follows that Eve can recover a message only
when she correctly intercepts entire sequence . Incorrect
detection on only one position results in (almost) random
decoded message.
III. ANALYSIS
An important step in studying protocol security is an
investigation of its robustness. Robustness of the protocol
informs how large disturbance is introduced by an
eavesdropper intercepting some information. QKD protocols
can be secure when they are partly or completely robust [4].
However, security requirements for QSDC protocols are much
more stringent because of the absence of privacy amplification
step. As a result partly robust QSDC protocols are considered
insecure and complete robustness guarantees only quasi
security, i.e. there is a finite, non-zero probability that
eavesdropper intercepts some part of the message without
being detected. Although similar property also holds for classic
cipher, the problem with QSDC security lies in fact, that
offered eavesdropping detection probability is relatively low
and Eve is detected with reasonable probability only after
sufficiently large number of protocol cycles. The pre- and post-
processing steps introduced herein provide all or nothing logic
in the message interception possibility. Thus proposed protocol
is insecure only when its quantum part is not robust. Contrary,
if the quantum part is partly robust or completely robust then
introduced classic pre- and post-processing steps provide
protocol computational security. In the following it will be
proven that quantum part is robust in lossless quantum
channels and partly robust in a lossy case. This renders that Eve
intercepts no useful information and protocol is secure. The
provided security margin is related to the quantum transmission
quality and QBER introduced by the eavesdropping.
Let us consider robustness of the protocol in the noiseless
quantum channel case. As it follows from the Stinesprings
dilation theorem, the most general quantum operation, which
may be performed on the signal qubit by an eavesdropping Eve
is described by an unitary operation that entangles the signal
qubit with the ancilla system of dimension
. Such a
transformation is described by four complex numbers
, (3)

(4)
where
denotes Eve's probe states and is the initial

state of the ancilla which is not entangled with the signal qubit.
The normalization ensures that
. The same
entangling operation has to be applied to the message and
control qubits because Eve acquires information what mode
has been used after the qubit has been stored in Bob's register.
Eve is detected with probability
and
when
legitimate users test in Z basis. It follows from (3) and (4) that
Vol. 2, No. 11, 2011
19 | P a g e

, (5)
where . Thus she is detected in X
basis with probability

(6)
because is unitary. Thus overall Eve's detectability is
minimized when
. But in that case Eve's

probe space is limited to two states and an entangling operation
may be reduced to simple CNOT in which signal qubit is used
as the control one. At the same time such operation provides
maximal information about the state of the signal qubit when
Eve performs measurements in Z basis. The quantum part of
the protocol is completely robust because Eve can't intercept
any information without risking to be detected. However,
robustness of the quantum transmission implies only quasi
security of the QSDC protocol.
Let us further consider how introduced preprocessing
improves the security characteristic of the protocol and assume
that legitimate users use ECC code able to recover from a given
QBER although they operate in noiseless channel. Such
assumption is favorable to the eavesdropper, as she can now
intercept some signal particles and her actions will be
undistinguishable from the noise. Because Eve is detected in X
with probability , it means that she can peek
of signal qubits per block without inducing an alarm. This is
the best case scenario for the eavesdropper. If the channel has
been noisy Eve would have to attack a less percentage of
particles to hide herself behind the total limit of errors. Thus it
may be assumed that she knows fraction of
and
sequences and the rest part of these sequences remains

random to her.
The preprocessing key is recovered as
.
But correctly recovered are only these fragments for which bits
on corresponding positions of sequences
and
are
correct and that happens with probability
. Thus to
recover the message
malicious Eve has to

attack a key space [
] of a well established
cipher and guess [ ] bits of the sequence
.
It follows that number of bits that have to be tested in brute
force attack exceeds for ( ) .
Available presently quantum channels may provide a better
performance. Thus computational complexity of an attack on
the protocol exceeds complexity of brute force guessing of the
message contents. Protocols with such property are regarded in
classic cryptography as secure. It is worth noting that although
computational complexity of the brute force attack has been
considered above, the proposed protocol does not require
establishment of the shared secret for secure operation.
Moreover, the block cipher used in the protocol works only in
encryption mode, thus may be replaced by another
cryptographic primitive providing randomization of the input
data, for instance, a stream cipher generator or keyed hash
function.
The performance of the described protocol is determined by
the overhead related to the transmission of test qubits and a
check block
. The number of test qubits must provide

reliable estimation of the channel quality within one data block
because decision about channel reliability should be taken
before the sequence
with encoded message is sent. As a

matter of fact the control protocol cycles may be disabled
during encoded message transmission as at this point of
protocol execution is too late for the eavesdropping detection.
The overhead related to transmission of the check sequence
may be diminished for messages longer than bits. In such
case message is padded and divided onto blocks
,
,
and each block is processed and sent independently
. The block with encoded preprocessing key
is sent as the first one.

IV. CONCLUSION
A single photon based protocol for quantum secure direct
communication has been proposed. Although its quantum part
is only quasi secure, the classic pre- and post-processing of the
message improves protocol security to the desired level. The
introduced protocol has very small demands on quantum
resources and can be, in principle, practically implemented in
the near future. Although the protocol is not unconditionally
secure, the provided security margin is high in noisy quantum
channels. It also offers some advantages compared to quantum
key agreement schemes proposed so far. The unconditional
security of QKD protocols has been proved in the limit of the
infinite length of the block being processed and the length of
the secret key is less than 50% of qubits sent. However,
efficiency of QKD scales very badly with the decrementation
of the sequence size and for moderate blocks of size
the
efficiency does not exceed 2% [17]. The proposed approach
offers improved efficiency at the price of the computational
security. It is also more versatile as it may be used for
confidential exchange of short sensitive messages without key
agreement and for regular QKD as well. Protocol also offers
also some advantages compared to presented so far QSDC
schemes [2,5] the quasi security limitation has been
conquered without requirement of large quantum memory
registers which are out of the reach of the present state of the
art technology.
REFERENCES
[1] N. Gisin, G. Ribordy, W. Tittel and H. Zbinden, Quantum
cryptography, Rev. Mod. Phys., vol. 74, pp. 145-195, 2002
[2] K. Bostrm and T. Felbinger, Deterministic secure direct
communication using entanglement, Phys. Rev. Lett., vol. 89, pp.
187902, 2002
[3] A. Beige, B. G. Englert, C. Kurtsiefer and H. Weinfurter, Secure
communication with a publicly known key, Act. Phys. Pol., vol. 101,
pp. 357-368, 2002
[4] M. Boyer, R. Gelles, D. Kenigsberg and T. Mor, Semiquantum key
distribution, Phys. Rev. A, vol. 79, pp. 032341, 2009
[5] M. Lucamarini and S. Mancini, Secure deterministic communication
without entanglement, Phys. Rev. Lett., vol. 94, pp. 140501, 2005
Vol. 2, No. 11, 2011
20 | P a g e
[6] K. Bostrm and T. Felbinger, On the security of the ping-pong
protocol, Phys. Lett. A, vol. 372, pp. 3953-3956, 2008
[7] G. L. Long, F. G. Deng, C. Wang, X. H. Li, K. Wen and W. Y. Wang,
Quantum secure direct communication and deterministic secure
quantum communication, Front. Phys. China, vol. 2, pp. 251-272, 2007
[8] F. G. Deng, G. L. Long and X. S. Liu, Two-step quantum direct
communication protocol using the Einstein-Podolsky-Rosen pair block,
Phys. Rev. A, vol. 68, pp. 042317, 2003
[9] S. Lin, Q. Y. Wen, F. Gao and F. C. Zhu, Quantum secure direct
-type entangled states, Phys. Rev. A, vol. 78, pp.
064304, 2008
[10] J. Wang, Q. Zhang and C. Tang, Quantum secure direct communication
without a pre-established secure quantum channel, Int. J. Quant. Inf.,
vol. 4, pp. 925934, 2006
[11] G. Gao, Two quantum dialogue protocols without information
leakage, Opt. Commun., vol. 283, pp. 2288-2293, 2010
[12] G. F. Shi, X. Q. Xi, M. L. Hu and R. H. Yue, Quantum secure dialogue
by using single photons, Opt. Commun., vol. 283, pp. 1984-1986, 2010
[13] F. G. Deng and G. L. Long, Reply to Comment on Secure direct
communication with a quantum one-time-pad, Phys. Rev. A, vol. 72,
pp. 016302, 2005
[14] D. Fu-Guo and L. Gui-Lu, Quantum privacy amplification for a
sequence of single qubits, Commun. Theor. Phys., vol. 46, pp. 443,
2006
[15] K. F. Reim, P. Michelberger, K. C. Lee, J. Nunn, N. K. Langford and I.
A. Walmsley, Single-photon-level quantum memory at room
temperature, Phys. Rev. Lett., vol. 107, pp. 053603, 2011
[16] R. L. Rivest, All-or-nothing encryption and the package transform,
LNCS, vol. 1297, pp. 210-218, 1997
[17] V. Scarani, H. Bechmann-Pascanucci, N. J. Cerf, M. Dusek, N.
Lutkenhaus, M. Peev, The security of practical quantum key
distribution, Phys. Rev. Lett., vol. 107, pp. 053603, 2011
AUTHORS PROFILE
P. Zawadzki received M.S. degree in theoretical physics from Silesian
University, Katowice, Poland, in 1989 and Ph.D. degree in electromagnetic
engineering from Silesian University of Technology, Gliwice, Poland, in
1998. His research interests include numerical simulation of interactions
between telecommunication infrastructure with the lightning electromagnetic
pulse, data protection in telecommunication networks and quantum
information processing in the context of quantum cryptography.

Vol. 2, No. 11, 2011
21 | P a g e
A Novel Implementation of RISI Controller
Employing Adaptive Clock Gating Technique

M.Kamaraju
Professor & Head, Dept.of ECE
Gudlavalleru Engineering College, Gudlavalleru, INDIA
Praveen V N Desu
M.Tech ES group, Dept of ECE
Gudlavalleru Engineering College, Gudlavalleru, INDIA

AbstractWith the scaling of technology and the need for higher
performance and more functionality power dissipation is
becoming a major issue for controller design. Interrupt based
programming is widely used for interfacing a processor with
peripherals. The proposed architecture implements a mechanism
which combines interrupt controller and RIS (Reduced
Instruction Set) CPU (Central processing unit) on a single die.
RISI Controller takes only one cycle for both interrupt request
generation and acknowledgement. The architecture have a
dynamic control unit which consists of a program flow controller,
interrupt controller and I/O controller. Adaptive clock gating
technique is used to reduce power consumption in the dynamic
control unit. The controller consumes a power of 174w@1MHz
and is implemented in verilog HDL using Xilinx platform
Keywords- Interrupt; Controller; Clock gating; power.
I. INTRODUCTION
The Interrupt Controller [1-2] is a device commonly found
in computer systems (both single-processor and
multiprocessors) which deals with interrupts generated by the
peripherals and the processors handle the interrupt priorities,
and delegates the execution to a processor.
The general purpose processors provide one or more
interrupt request pins that allows external devices to request
the service provide by CPU. Consider a case in which
processor can handle a large number of interrupts which are
come from external devices. The design requires a separate
interrupt controller which is interfaced to the processor. This
increases the complexity of design. More over the processor
needs some extra- interfacing Circuits which decreases the
performance and increase the power consumption of the
overall system. The proposed architecture combines the
interrupt controller and RIS CPU employs an adaptive clock
gating to reduce the overall power consumption.
Power [18] is the one of the design constraint, which is not
only applied to portable computers and mobile communication
devices but also for high-end systems. Power dissipation
becomes a bottleneck for future technologies.
In the early days designers treat the clock signal should not
be disabled or disturbed. But clock signal is a major source for
power dissipation and it is a dynamic in nature because clock
signal is feed into several blocks in the processor. Because all
the blocks usage varies within and across a processor, all the
blocks not used all the time and gives a chance to reduce the
power consumption of unused blocks. Clock gating is an
efficient technique to reduce the dynamic power dissipation.
By anding the clock signal with gated control signal clock
gating technique disables the clock signal to the block when
the block is unused. Adaptive clock gating technique [4-6] is
one of thetechnique used to reduce the dynamic power of the
clock. In this technique clock gating enable signal generated
by the block itself depending upon the usage and this
technique will reduce the burden on the control unit for
generating clock gating signal.
Interrupt handling mechanism [8] provides how the
interrupt is handled by processor. There are various clock
gating techniques [4][5] to reduce the dynamic clock power
dissipation. The interrupt controller [1][2] takes two cycles
(one cycle for generating the interrupt request and another
cycle for Acknowledgement) to process the interrupt. RISI
Controller takes only one cycle for both interrupt request
generation and acknowledgement. The following section
provides a brief overview of architecture of the RISI
CONTROLLER and explanation about the implementation
and hardware consideration. Along with a brief description
about each block present in the architecture is given. Finally a
few notes on simulation results
II. ARCHITECTURE OF RISI CONTROLLER
The reduced instruction set interrupt controller (RISI
Controller) architecture mainly consists of ALU (Arithmetic
and logical Unit), Port Controller, Interrupt controller and
Register Array and its block diagram is shown in the Fig. 1. It
contains RISCPU, Interrupt controller, Port controller and
Program flow controller. These blocks are connected by
internal buses.
Figure 1. Block diagram of a RISI Controller
The internal architecture of RISI Controller is shown in the
figure2.
Vol. 2, No. 11, 2011
22 | P a g e

Figure 2.Internal architecture of RISI CONTROLLER
The instruction length of RISI Controller is 16-bit
wide.RISI Controller has three flags namely carry, zero and
interrupt flags. Both zero and carry flags are affected only
during the execution of arithmetic and logical instructions and
these are also useful for determine the flow of execution when
branch and jump instructions take place. CPU checks the
interrupt flag after completion of every instruction to know
whether interrupt is available or not.ALU is capable of
performing the Arithmetic (Like Addition and subtraction) and
Logical operations (like And, Or, Xor and Cmpl). There are no
special purpose registers in the CPU like accumulator and
there is no priority among them.
RISI Controller has multi read port and single write port.
Generally read operations are performing during the positive
edge of the clock and write operation is performing during
negative edge of the clock. Stack is used to store up to four
addresses during interrupt and Branch related instructions.
Port controllers take care of the read and write operation. An
8-bit address value provided on the PORT bus together with a
READ or WRITE strobe signal indicates the accessed port.
The port address is either supplied in the program as an
absolute value or specified indirectly as the contents of any of
the eight registers
There are some specific instructions useful for the
controlling of interrupt controller present in the RISI
CONTROLLER.

III. INTERRUPT CONTROLLER
Modern CPUs [15, 16] provide one or more interrupt
request pins that allows external devices to request the service
provide by CPU. Interrupt controller are used to increase the
number of interrupt inputs available to CPU. The block
diagram, of interrupt controller is shown in figure3.
Figure3.Block diagram of Interrupt Controller
Interrupt controller composed with three blocks. They are
Interrupt Register block, Edge interrupt detection unit and
Interruptrequest generation unit. Interrupts are identified by
interrupt detection unit during the negative clock edge of the
clock. Whenever interrupts are detected, check for the
corresponding interrupt input masked or not. Unmasked
interrupt input set the corresponding bit in the interrupt status
register. IRQ generation unit generates the interrupt request by
using the IVR contents. Interrupt request reaches the CPU
send an acknowledgement signal.
Int_inputs are used to monitor the interrupts coming from
various peripherals or external devices. Each interrupt register
has a unique address and identified by using Addr input. To
write the contents of Data input into the interrupt registers
require a high valid_wr input. Intr flag input indicates the
status of interrupt flag present in the CPU.
Interrupt Detection Unit detects the interrupts coming from
peripheral or external devices and actives the logic to generate
enable interrupt to controller. It monitors the interrupt inputs
composed of interrupt signal coming from external devices or
peripherals and rises enabled interrupts according to arrival
signals, Interrupt Request (IRQ) Generation unit contains the
Generation logic of the Interrupts towards the
processor.Interrupt requests generation is also configurable as
either a pulse output for an edge sensitive request or as a level
output that is cleared when the interrupt is acknowledged.
Interrupt Registers handles the interrupt priorities, deciding
which, interrupts are enabled or disabled and managing of
interrupt acknowledgements. It contains the following
Registers
Interrupt Status Register (ISR) indicates which interrupts
are active and the format is shown in the figure 4. All bits in
the ISR are set to zero default. Any bits are set to 1 indicates
that the corresponding interrupt is active otherwise no active
interrupts are available
Vol. 2, No. 11, 2011
23 | P a g e

Figure 4. Interrupt Status Register
Interrupt Pending Register (IPR) gives the information
about the interrupts that are both active and enabled. By
default all the bits in IPR are set to zero. Any bit set to 1
indicates that the corresponding interrupt is waiting for
processing and 0 indicate no interrupt isavailable. The IPR is
shown in the figure 5

Figure 5. Interrupt Pending Register
Interrupt Enable register (IER) keeps track which
interrupts are allowed to be handled. Writing a 1 to a bit in
this register enable the corresponding interrupt input signal.
Writing a 0 to a bit disable or mask the corresponding
interrupt input signal
Interrupt Acknowledge Register (IAR)is used to disabling
the interrupt request with selected interrupt input. Writing a
1 to a bit location will clear the interrupt request that was
generated by the corresponding interrupt input and Writing 0
does nothing. The IAR is shown in the figure 6

Figure 6. Interrupt Acknowledge Register
Interrupt Vector Register (IVR) contains the ordinal value
of the highest priority, enabled, active interrupt input. INT0
(always the LSB)is the highest priority interrupt inputand each
successive input to the left has a correspondingly lower
interrupt priority. If no interrupt inputs are active then the IVR
will contain all ones. The Interrupt Vector Register (IVR) is
shown in the figure 7.

Figure 7.Interrupt Vector Register
Master Enable Register (MER) is used to enabling the
interrupt requests to the processor. Writing a 1 to ME bit
enables the IRQ output signal and 0 to ME bit disable the
IRQ output signal in other words masking all the interrupt
inputs. The Master Enable Register (MER) is shown in the
figure 8.

Figure 8.Master Enable Register
The following are the important features of Interrupt
controller present in the RISIController. They are i) Priority
between interrupt requests is determined by vector position.
The least Significant Bit (LSB, bit 0) has the highest priority.
ii) Interrupt enable register for selectively disabling or
enabling of individual interrupt inputs. iii) Master enables
register for disabling interrupt request. iv) Easily cascaded
to provide additional interrupt inputs. v) Low power and less
area
IV. IMPLEMENTATION
The RISI CONTROLLER is implemented using Xilinx
platform on Virtex4 FPGA Family in VerilogHDL. The
flowchart for the Interrupt controller is shown in the figure 9.
Application specific instructions [13] were designed for
controlling the interrupt controller along with general purpose
instruction set.
By using Application specific instructions [13] processor
can perform several operations on interrupt controller. The
operations like masking or unmasking of interrupts disable
interrupt request for the execution of important instructions.
Interrupt controller uses fixed priority algorithm for generating
the interrupt request.
The flow chart for the interrupt controller is shown in the
figure 9 and is explained below. Various peripherals or an
external device wants the services provided by CPU. They
generate an interrupts to interrupt controller. Interrupt coming
from various external devices are identified by edge interrupt
detection unit during the negative half cycle of the clock. Edge
detection unit in the interrupt controller check whether the
interrupts are masked or not. This information is present in the
IER register.

Figure 9. Flow chart for interrupt controller
Vol. 2, No. 11, 2011
24 | P a g e
Enable interrupts are obtained by anding the
Interrupt_inputs with the contents of IER register. This
information is stored in the ISR register.
The MER must be programmed based on the intended use
of the Interrupt Controller.There are two bits in the MER: the
Hardware Interrupt Enable (HIE) and the Master IRQ Enable
(ME). The ME bit must be set to enable the interrupt request
output.Check whether the Msb bit (ME) in MER register is set
or not. If ME bit in the MER register is not set the interrupt
requests are not generated. Check the status of the interrupt
flag. If the flag bit is set Interrupt request is generated and wait
for acknowledgement which is coming from the processor.
Next interrupt request is generated only when the
acknowledgement is received otherwise; wait until the
acknowledgement is received.
By the execution of the following code the RISI Controller
disables the interrupt request and enables it again after some
time. The simulation results regarding the execution of the
program1 are shown in the results section
Program 1

MID 00 ----- Disables the interrupt Request
MOV R4, 5F ----- Load the immediate data5F into register R4
MOV R5, 5F ----- Load the immediate data 5F into register R5
ADD R4, R5 -----Add the contents of R5with R4 and result is
Stored in R4
MIE 80 ----- Enables the interruptrequest

RISIController is capable to handle 8 interrupt inputs at a
time. Interrupt controller uses positive clock cycle for interrupt
request generation whereas negative clock cycle used for
receiving an interrupt Acknowledgement which is coming
from CPU.
V. RESULTS
All the modules of RISI CONTROLLER are simulated and
verified using the Xilinx tools. The simulation result of
Interrupt controller shown in figure 10. When the reset is high
all the register present in the interrupt controller are loaded
with default values. Interrupts are identified during the
negative edge of the clock signal. The content of IER indicates
whether the interrupts are mask or unmask. En_int is obtained
by anding the content of IER with Int_inputs.
Each register present in the interrupt controller have a
unique address. Registers are identified by using addr input.
The content ofEn_int is also stored in the ISR. Interrupt
request is generated only if both the ME and Intr flag arehigh.
Priority between the interrupts is generated by using the fixed
priority scheme. The contents of Clr_isr are set at the time of
interrupt request generation.
Top level timing diagram between interrupt controller and
RISCPU is shown in figure 11. Whenever the CPU receives
interrupt request, wait for the completion of current instruction
and store the address of next instruction in the stack. Load the
contents of program counter with address FF and send an
acknowledgement to the Interrupt Controller by activating the
addrin and valid_wr output.

Figure 10.Top level timing diagram of Interrupt Controller
Status of intr flag is changed from high to low when the
interrupt request is received. CPU does not receive any
interrupt even though Int_inputs are active because Me output
is low indicate that interrupt request is disabled. The status of
Meoutput is high then only the CPU receives the interrupt
request and it is shown in the figure 11.

Figure 11: Top level timing Diagram between the Interrupt controller and RIS
CPU
The top level module with input and out signals of the
RISI Controller and Interrupt Controller is shown in figures
12, 13 and the RTL schematic of the Interrupt Controller is
shown in the figure 14.

Figure 12. Signals of RISI CONTROLLER
Vol. 2, No. 11, 2011
25 | P a g e

Figure 13. Signals of Interrupt Controller(IC)
The RTL schematic of the Interrupt Controller of RISI
Controller is shown in figure 14. The Interrupt controller
consists of a Register block, edge interrupt detection unit and
IRQ generation unit. Interconnection between the blocks and
the input output signals of the Interrupt Controller module are
also shown in the figure 14.

Figure 14.RTL Schematic of Interrupt controller of RISI Controller
Figures 15 and 16 represents the layout of the RISI
Controller and Interrupt controller implemented onto the
Vertex4 FPGA family. In the Fig. 15 and 16 the colored area
represents the components of the RISI CONTROLLER that
are placed on the FPGA and the place and routed diagrams of
the Interrupt Controller are shown in figures 17 and 18
respectively

Figure 15. RISI Controller Layout on FPGA
Placement involves deciding where to place all electronic
components, circuitry, and logic elements in limited amount of
space. This is followed by routing, which decides the exact
design of all the wires needed to connect the placed
components.

Figure 16.Interrupt controller(IC) layout on FPGA
This must implement all the desired connections by
following the rules and limitations .The complete placed and
routed diagram of the RISI Controller is shown in figure 17.

Figure 17.Place and Route diagram of RISI CONTROLLER
The place and route diagram of theInterrupt controllerof
RISI Controller is shown in figure 18.

Figure 18. Place and Route diagram of Interrupt Controller(IC)
In the Table I the comparison of different parameters
related to the area of RISI Controller and Mpcore were
specified. The graph clearly shows that the number of slices
and LUTs utilized by the RISI Controller less than that of the
Mpcore. The graph also indicates in othersense that the area
occupied by RISI Controller on the FPGA is less.
TABLE I.COMPARISON OF DEVICE UTILIZATION PARAMETERS OF MPCORE AND
RISI CONTROLLER

Parameter Mpcore RISI
CONTROLLER
Number of slice
Registers
349 318
Slices containing
related logic
496 467
Gate count 7360 7162

Vol. 2, No. 11, 2011
26 | P a g e
The comparison of different parameters related to RISI
Controller signal delays were shown in the figure 19. The Net
Skew for clock is the difference between the minimum and
maximum routing only delays for the clock.

Figure 19.Comparison of delay parameters for Mpcore and RISI Controller
Power is one of the main constrain in the design of
controller. Clock signal [4] contribute major power
consumption source. To reduce the clock power
dissipationclock gating techniques are employed. The power
consumption of the RISI CONTROLLER was described in the
figures 20 and 21. The figures also give us a comparative view
of the power consumption of the Mpcore and RISI
CONTROLLER cores.
In the figure 21 the deviation of the curves indicate that the
operating voltage increases the power consumption of the
Mpcore increases more rapidly when compared with designed
controller power. Similarly figure 20 describes the variation of
power with respect to frequency for the Mpcore and RISI
CONTROLLER.

Figure 20. Power consumption of RISI CONTROLLER at 1.2v

Figure 21.Plot b/w Voltage and Power of Mpcore and RISI Controller
at 1 GHz
The Power-delay product is simply the product of the
power consumption and the time delay. The smaller value of
the power-delay product, performance of the design is better.
Since this RISI CONTROLLER has almost negligible power-
delay product value, it indeed has a better performance in
terms of the speed and power dissipation.
TABLE II.COMPARISON OF MPCORE AND RISI CONTROLLER

Table II shows the comparison of different parameters
related to the Mpcore and RISI CONTROLLER; it has better
performance than the Mpcore.
VI. CONCLUSION
As the amount of data transferred between the main
processing unit and peripheral devices increases, the frequency
of interrupts from peripheral devices also increases. The RISI
controller designed integrates a RISCPU and Interrupt
controller onto the single chip. The advantage of the designed
chip is, it can handle an interrupt fast and effectively. It
occupies less area and consumes less power. More over an
integrated CPU in the design performs the necessary
operations related to interrupt controller apart from the regular
operation. The scope for increasing the number of interrupts
up to databus width isprovided in the design and also extended
to multiprocessor.

Vol. 2, No. 11, 2011
27 | P a g e
REFERENCES
[1] A. Tumeo, M. Branca, L. Camerini, M. Monchiero,G. Palermo,F.
Ferrandi and D. Sciuto, An Interrupt Controller for FPGA-based
Multiprocessors, International Conference on Embedded Computer
Systems: Architures, Modeling and Simulation, July 2007,pp. 82-87.
[2] Wei Chipni, Li Ahao lin, Zheng Qingwei,Ye Jianfei, and Li Shenglong,
Design of a configurable multichannel interrupt controller,Second
Pacific-Asia Conference on Circuits,Communications and System, vol.
1,Aug 2010, pp. 327330.
[3] Pizhou Ye,and Chaodong Ling, A RISC CPU IP core,Second
International Conference on Anti-counterfeiting,Security and
Identificatiom ,August(2008), pp. 356359,2008.
[4] H. Jacobson,P. Bose, Zhigang Hu,A. Buyuktosunoglu, V.Zyuban,R.
Eickemeyer,L. Eisen,J. Griswell,D. Logan,Balaram Sinharoy, and J.
Tendler, Strenching the limits of clock-gating efficiency in server-class
processors,Eleventh International Symposium on High-performance
Computer Architecture,Feburary (2005), pp . 238- 242,2005.
[5] Xiaotao Chang, Mingming Zhang, Ge Zhang, Zhimni Zhang and Jim
Wang,Adaptive Clock Gating Technique for Low Power IP Core in
SoC Design,IEEE International Symposium on Circuits And
Systems,May(2007),pp. 2120 2123, 2007..
[6] S. Ghosh, D. Mohapatra, G. Karakonstantis and K. Roy, Voltage
Scalable High-speed Robust Hybrid Artihmetic Units Using Adative
ClockingIEEE Transactions on Very Large Scale Integration
Systems,September(2010),Vol. 18, pp.1301- 1309,2010.
[7] Hai Li, S. Bhunia,Y. Chen,T.N. Vijaykumar and K.Roy, Deterministic
clock gating for microprocessor power reduction, Ninth International
Sympossium on High-Performance Computer Architecture, February
(2003), pp. 113 122, 2003.
[8] E. Ozer, S.W. Sathaye, K.N. Menezes, S. Banerjia, M.D. Jenningsand
T.m. Conte, A fast interrupt handling scheme for VLIW
processors,International Conference on Parallel Architecures and
Compilation Techniques,October(1998),pp. 136 -141 ,1998.
[9] G. Kane and J. Heinrich, MIPS RISC Architecture: reference for the
R2000,R3000,R6000 and the new R4000 Instruction set computer
Architecture, Prentice-Hall,Englewood Cliffs,NJ,1992.
[10] Krall. A, An extended Prolog instruction set for RISC processors in
VLSI for Artifical Intelligence and Neural Networks,J.G. Delgado-Frias
and W.R. Moore, Eds-Plenum Press, New York,pp. 101-108,1991.
[11] Motorola Inc.M6800 8/16/32- Bit MICROPROCESSOR,avilable
athttp://www.freescale.com/files/32bit/doc/ref_manual/MC68000UM.pd
f
[12] J. Rose,A.E. Gamaland A. Sangiovanni and A. Vincentelli,
Architecture of Field-Programmable Gate Arrays, Proc. IEEE,Vol81,no
7,July,pp. 1013-1020,1993.
[13] J. Vanprate, Gossens .G, D. Lanner, and H. De man, Instruction set
definition and insstruction set selection,in Proc. Of seventh
International Symposium on High-Level Synthesis,pp. 11- 16,1994.
[14] ARM11 Mpcore available athttp://www.arm.com.
[15] IBM. Multiprocessor Interrupt Controller Data Book,March,2006.
[16] Xilinx OPB Interrupt Controller (v1.00c),January 2005.
[17] A. De Gloria,Paolo Faraboschi and Mauro Olivieri, A Self
TimedInterrupt Controller: ACase study in Asynchronous
Micro_Architecture Design Seventh Annual IEEE International ASIC
Conference and Exhibit ,pp. 296 299,1994.
[18] G. Palumbo,F. Pappalardo,and S. Sannella Evaluation on power
reduction applying gated clok approes,IEEE International Symposium
on Circuita and Systems,Vol.4,pp.IV-85-IV-88,2002.
[19] Nakasimha .K,S. Kusakabe,H. Taniguchi,and M. Amamiya Design and
implementaion of Interrupt packaging mechanisum,International
Workshop on Innovative Architecture for Future Generation High-
performance processors and systems, pp. 95-102,2002
[20] J.E. Smith, and A.R. Pleszkun, Implementing Precise Interrupts in
Pipelined Processors IEEE Trans. On Comp.,Vol.37,pp. 291-299,1975.
[21] Horelick Dale simple Versatile Camac Crate Controller and Interrupt
Priority Encoding Module , IEEE Transaction on Nuclear Science,Vol.
22,pp. 517- 520,1975.
[22] Qiurong Wang An Interrupt Management Scheme Based on
Application in Embedded system,Multimedia and Informetion
Technology,pp. 449-452,2008.
[23] Kamaraju.M, Lal Kishore.K, Tilak.A.V.N A Novel Implementation of
Application Specific Instruction-set Processor (ASIP) using Verilog
WASET Issue 59, NOV 2011.

Vol. 2, No. 11, 2011
28 | P a g e
Strength of Quick Response Barcodes and Design of
Secure Data Sharing System
Sona Kaushik
Student of Masters in Technology,
Birla Institute of Technology,
Mesra, Ranchi, India

Abstract - With the vast introduction of the wireless world, the
exchanged information now is more prone to security attacks
than ever. Barcodes are the information careers in the form of an
image. Their various applications have been discussed in brief
and also the structure, symbology and properties of barcodes.
This paper aims to provide an approach which can share high
security information over the network using QR barcodes. QR
barcodes are explained in detail to get the rigged understanding
on quick response technology. The design of data security model
to share the data over the network is explained. This model aims
to enable secure data share over the network. This model is a
layered architecture and protects the data by transforming the
structure of content. So barcodes are used to put tricks over the
information rather than directly using it for its noble
functionality.
Keywords-QR Barcode; Quick response technology; Data Security;
Information Security; Image processing; Data Sharing
Architecture
I. INTRODUCTION
The emergence of diverse networked data sources has
created new opportunities for the sharing and exchange of data.
The use of information has become a persistent part of our
daily life. Employees use information to perform elementary
job functions; managers require significant amounts of it for
planning, organizing and controlling; corporations leverage it
for strategic advantage.
Since the application of computers in administrative
information processing is also very important, computers have
become a key instrument in the development of information
processing. The rapid development of information technology
(IT) has helped to firmly establish the general attitude that
information systems are a powerful instrument for solving
problems.
Utilizing these emerging technologies, however, is not
without problems. People start considering their sensitive
information when it is transmitted through open networks; they
began worrying about using forged information for business;
and corporations worry about customer and investor confidence
if they fail to protect sensitive information. Protecting sensitive
information has consequently become a top priority for
organizations of all sizes.
The majority of existing systems focus on performance and
precision of data retrieval and information management. A
number of techniques are employed to protect information;
however, in many cases, these techniques are proving
inadequate. For example, while several information systems
use the add-ons security features to provide information
confidentiality (which allow users to share information from a
data media while keeping their channel private), these security
measures are insufficient.
Considering barcodes as an effective media to share
information, at present, the black and white two dimensional
barcode technology has developed more mature. American
National Standards Institute (ANSI) developed the international
standards of two-dimensional bar code, QR codes [1][2].
However, with the urge of increase in information, expanding
the field of bar code applications is thought to be good idea.
Barcodes are used to store the high capacity information in less
space and thus, stands as a good ideate.
In the upcoming section, the structure, information capacity
and other important properties of QR barcodes are explained.
In section 3, a model is proposed with an algorithm. Finally in
the last section, future work and conclusion for the model is
proposed.
II. BARCODE APPLICATIONS
An effective and innovative call from Denso introduced the
world with Barcodes which are so efficient in their work,
especially in specific domains like automated storage systems
and data hiding [3][4].
Barcodes finds good applications in inventory control
systems like stock level maintenance, control on incoming
goods and raw materials, etc. Also barcodes has production
control applications like status on information of
authentications of goods. They are also extensively used in
industries and business sectors related to automotive, food
processing, electronics, insurance, pharmaceutical, and so on
[5].
As if now, author does not find barcodes to be used to
secure data while transferring over the network. Before
knowing the proposed work using barcodes, how barcodes are
invented and what is the structure behind the creation of
barcodes is discussed in next section.
III. RESEARCH ON QR BARCODES
The strength of QR Barcodes can be identified as they
contain the forte of PDF 417 for its high data capacity, Data
Matrix for its reduce space printing and MAXI Code for its
high speed reading as presented in fig. 1.
Vol. 2, No. 11, 2011
29 | P a g e

Fig 1. The Development of QR Code.

A. Generating Codes
QR codes can be generated online with various code sizes.
For optimum readability we recommend a quiet zone area of 4
modules around the code. A module is the smallest pixel
element of a QR Code. For most phones a smaller quiet zone of
2 modules will work, but 4 modules are required in the official
documentation [8]. So we also recommend the module size to
be 4 so that it captures the most information within and
maximizes the capacity size. See fig. 2.
B. Encoding Process
There are three steps in the encoding process of colour bar
code: encoding the data information, generating error
correcting code, and generating symbol process [9]. Firstly,
according to the data information encoding mode, the data is
transformed into data stream. Meanwhile, the information of
length of data stream is added into the head of data stream.
Then each four bits in the data stream form a code and the
corresponding error-correcting code is generated through Reed-
Solomon algorithm. The error-correcting code is added to the
end of the data stream. Finally, all of the data codes will be
transformed into symbols and the function images are added
too as shown in fig. 3.

Fig. 2. QR Barcode Module.
C. Information Capacity
In the black and white two dimensional bar encoded
method, the data is transformed into binary data stream [10].
Then these binary data value, 0 and 1, are presented with black
and white symbols. Each module which can present 0 or 1, can
express two kinds of different information. So in the black-and-
white two-dimensional barcode with n symbol modules, the
information capacity will be 2n.

Fig. 3. The Encoding Process.

D. QR Barcode Structure
The QR barcode [6][7] is a two dimensional symbol
developed by Denso Wave in 1994. The code contains
information in both the x-axis and y-axis, whereas traditional
barcodes contain data in one direction only. The main structure
of the QR barcode is shown in fig. 4. The outer range is the
quiet zone. The upper-left, upper-right, and left bottom square
areas are used for position detection and pattern separators for
positioning.
There are six smaller squares which are the alignment
patterns. Additionally, the main area, which is colored grey, is
the kernel area of data and error correction code. The QR
codes size is decided by determining a symbol version based
on data capacity, character type, numeric, alphanumeric, etc.,
and error correction level, and by setting a module size based
on the target printers or scanners performance level. The
placement of finder patterns and timing patterns can be seen in
fig. 5.
Vol. 2, No. 11, 2011
30 | P a g e

Fig. 4. The Structure of QR Barcodes.

Fig. 5. The Finder and Timing Pattern of QR Barcode.

E. QR Code Structure in Detail
Fig. 6 shows an example of QR code symbol. The figure is
version 1 (type 2) and the module is 21 X 21 cells, vertical 21
cells and horizontal 21 cells. This version is specified from 1 to
40, increased by 4 cells per one version up. The maximum
version is 40 and the size is 177 X 177 modules. Fig. 6 is a case
of the QR code version 1 modules that are arranged in a grid
pattern of black and white squares. In this QR code symbol
[11][12][13], there are three position detection patterns (Finder
patterns) in the upper left corner, bottom left and top right
corner. Then the timing pattern is placed between every one of
these position detection patterns. Additionally, alignment
patterns are introduced in the version 7 or higher. Then Table I
shows the main specifications of the QR code. There are four
modes available,
(1) number mode,
(2) alphanumeric mode,
(3) 8 bit byte mode and
(4) kanji and kana characters mode.
Vol. 2, No. 11, 2011
31 | P a g e

Fig. 6. Example Structure of QR Code.

It can also be combination of these modes. The RS (Reed
Solomon) error correction code is used for recovering from
symbol dirty or transmission error. There are four levels of
error correction capability, Level L: about 7% recovery,
Level M: about 15% error recovery, Level Q: about 20%
error recovery and Level H: about 30% recovery.
IV. PROPOSED ALGORITHM
An algorithm is proposed here which secures the
information while being shared on the network. The idea is
to create a random number corresponding to the information
saved in the database. The random number behaves as the
key to the information in the database. Each random unique
key point to the unique information placed in the database
server.
Barcodes are used for the obvious reasons. They are
acceptable worldwide. Various barcode readers are openly
available on the internet and barcodes can be widely scanned
in mobile devices.
For the random number created initially, a quick
response barcode is created. QR barcode generators are
widely available in the market. The preferable size of QR
barcode in this algorithm is 800 length and 800 height to
utilize more information space in barcodes.
The barcode created is now used to create two identical
looking barcodes with minor differences in each which are
not detectable by naked eyes. The identical image one is
created by first detecting the black continuous series of pixel
in barcode. Then the first black pixel from the series is
formatted to white or base color. The process is repeated for
all such black series of pixels in the image. This image is
now named as the False Image One. False image one, on
completion of formatting for each black series of pixel in the
image, then looks identical to the original image.
TABLE I. MINIMUM QR CODE SPECIFICATIONS.
Item Specifications
Error
Correcting
Code
RS Code Data
BCH Code
Format Information
Version Information
Characters
Number
10 bit coding per 3 number
digits
Alphanumeric 11 bit coding per 2 characters
8 bit byte 8 bit coding
Kanji 13 bit coding per 2 characters
Version
1 21 x 21 modules
2 25 x 25 modules
40 177 x 177 modules
Error
Correcting
Level
L about 7%
M about 15%
Q about 25%
H about 30%
Finder
Pattern
1:1:3:1:1
3 co-centric squares
7x7, 5x5, 3x3 modules
Alignment
pattern
1:1:1:1:1
3 co-centric squares
higher version 2
5x5, 3x3, 1x1 modules

Now the second false image is created out of the original
image. Same process as in creating false image one will be
followed with a small difference. Black series of pixels are
identified and then unlike false image one, this time the last
black pixel from the series is formatted to white or base color
instead of first black pixel in the series.
This is another identical looking image to the original
barcode created and thus, called False Image Two.
As a result of steps described above, we get three images all
identical to each other. But only one image which is original
can be read by the Barcode Readers.
The false image one is then sent over the network to the
destination. After a predefined scheduled time difference, false
image two is also sent to the destination. The two images are
scanned at the receivers end and fed to the algorithm decoder.
SUMMARIZED ALGORITHM: (SEE FIG. 7)
Step 1: Create a unique alpha numeric random number of
length 1024 characters corresponding to the message to be
secured.
This will allow 62
1024
different combinations that are
infinite.
Step 2: Of this random number, a QR Barcode (Quick
Response) Image is generated of size 800 * 800.
Vol. 2, No. 11, 2011
32 | P a g e

Fig. 7. Proposed Algorithm.

Step 3: Two Barcode images are created from the original
barcode image each of size 800*800. One image will be
manipulated by inverting one left most black pixel of each
black continuous series of black pixels. Likewise, other image
will be manipulated by inverting one right most black pixel of
each black continuous series of black pixels. This will give the
effect of three similar images.
Step 4: These two images are sent to the destination end
with the time difference of 3 to 7 minutes (randomly created,
on the fly) and the reverse process is followed at the
destination to get the secured message.
V. CONCLUSION
Protection of sensitive information is a growing concern
around the globe. Securing critical data in sectors like business,
healthcare and military, has become the first priority of
sensitive information management. Failing to protect this asset
results in high costs, loss of customers and investor confidence
and even threaten national security.
QR Codes are two dimensional barcodes that can contain
any alphanumeric text. Quick Response (QR) codes are
versatile. A piece of long multilingual text, a linked URL, an
automated SMS message, a business card or just about any
information can be embedded into the two-dimensional
barcode. Barcodes are used to store the high capacity
information in less space and thus, stands as a good ideate.
An algorithm is proposed in this paper which efficiently
transfers the data with high security with the aim of able to
protect sensitive information. This can be used for very high
security data and its complexity makes it more efficient to
secure the information. Thus, a competent and innovative way
of utilizing the barcode technology is proposed. The system is
efficient enough to handle the 1024 characters and leads to
completely efficient and reliable results. Image processing
makes the systems processing more complex but stands for its
application purpose.
VI. FURTHER WORK IN PROGRESS
The implementation of the design of secure data exchange
system is in progress. Also a real time application case study is
in progress to its accomplishment. Also the capacity of the
barcode can be increased by using color barcode.
REFERENCES
[1] International Organization for Standardization: Information Technology
International Symbology Speciation Data Matrix [S]. ISO/IEC16022
(P), 2000.
[2] Pavlidis T, Swartz J. Fundamentals of barcode information theory[J].
IEEE Computer, 990, 23(4): 74- 86.
[3] Thota Sriram, K.V.Rao, S Biswas, Basheer Ahmed, Application of
barcode technology in automated storage and retrieval systems, BHEL.
[4] Bing Nan, Ming-Chui-Dong Y Mang, From Codabar to ISBT 128:
Imlementing Barcode Technology in Blood Banking Automation
System, Proceedings of the 2005 IEEEEngineer in Medicine and
Biology 27
th
Annual conference, Shanghai, China, 2005.
[5] Barcoding : A complete tracking system from design to despatch; Mark
Marriot, Numeric Arts Ltd., (Logistics today).
[6] Wen-Yuan Chen, Jing-Wein Wang; Nested image steganography
scheme using QR-barcode technique Optical Engineering 48_5, 057004
_May 2009_Vol. 48(5).
[7] See http://www.denso-wave.com/qrcode/qrfeature-e.html_.
[8] A Guide to Kaywas QR code solutions, See
http://mobile.kaywa.com/qr-code-data-matrix.
[9] Wicker, Bhargava. Reed - Solomon Codes and Applications [J]. IEEE
Press, 1994.
[10] Zhi Liu, Herong Zheng, Huaguo Jia; Design and Implementation of
Color Two-Dimension Barcode with High Compression Ratio for
Chinese Characters, International Conference on Information
Engineering and Computer Science, 2009 (ICIECS 2009).
[11] Wakahara, Toshihiko; Yamamoto, Noriyasu; Image Processing of 2-
Dimensional Barcode, Conference on Network-Based Information
Systems (NBiS), 2011 14th International .
[12] Japanese Industrial Standards, Two Dimensional Symbol-QR-Code-
Basic Specification JIS X 0510, October 2004.
[13] T. J. Soo, QR Code, Synthesis Journal, pp..59-78 2008.
[14] Xianping Wu; Security Architecture for Sensitive Information
Systems.
AUTHORS PROFILE

Sona Kaushik belongs to National Capital Region, New
Delhi, India. She received the B.Tech. degree in
Information Technology in 2007. She is currently working
as a System Engineer in a reputed IT Organisation and
pursuing Masters in Technology from Birla Institute of
Technology, Mesra, Ranchi, India. Her research interests
are in information security, network security, security
engineering and cryptography.
Vol. 2, No. 11, 2011
33 | P a g e
Graphing emotional patterns by dilation of the iris in
video sequences

Rodolfo Romero Herrera
Departamento de Posgrado, Escuela
Superior de Computo (ESCOM -
IPN)
Mxico D.F.
Francisco Gallegos Funes
Escuela Superior de Ingeniera
Mecnica y Electrica
(ESIME IPN)
Mxico D.F.
Saul De La O Torres
Departamento de Posgrado, Escuela
Superior de Computo (ESCOM -
IPN)
Mxico D.F.

Abstract For this paper, we took videos of iris of people while
induced a feeling of joy or sadness, using videos to motivate the
states affective. The manuscript implemented is a system of
recognition affective pattern by dilating the iris, with which
extracted images of the videos. The results obtained are plotted to
facilitate the interpretation. A suitable treatment occurred for
locating the pupil and the obtaining of the diameter of the pupil.
The graphics are based on statistical time intervals.
Within the research found that the iris diameter varies depending
on your mood present in the subject of study. There is software
that can detect changes in the pupil diameter, however it is also
develops software, the main objective is to detect changes with
respect to affective states and in this study is the main
contribution. The joy and sadness were the emotional states that
may differ. The system presents graphs that can be observed
when analyzing the dependence between feelings and dilation
present in the eye.
Keywords- Pupil dilation; joy; sadness; mathematical morphology;
average interpolation.
I. INTRODUCTION
The emotion recognition is a step to truly intelligent
machines, therefore is essential, for recognition the emotional
pattern recognition to understand the human intelligence [1].
Based on the growing interest in using the technology to
read emotions in a human (affective systems) [2] [3] [4] and
the fact that they have found evidence that emotions cause
changes in the iris [5], The system developed is a system to
recognize and classify the patterns that occur in the iris of a
person while changes your mood, focusing in two of the basic
human emotions: joy and sadness. Although the reader should
know that Paul Ekman bases their work on six basic emotions
[6].
II. METHODOLOGY
The pupil is a hole that is images of black. In order to
eliminate the illumination and brightness was used HSV color
format [7][8], the brightness component is greater than 0.5, as
it goes from 0 to 1; however it is worth taking a higher than
average brightness. Located these points, it creates a new pixel
with the same value H and S, but lower in V. MAX is the
maximum value of the components (R, G, B), and MIN the
minimum value, then we have that HSV by (1) (2) and (3)
allow the conversion to this format [9].

{

V = MAX
For the analysis of the eye, it is isolating the darkest
elements of the image, which also are tabs and areas outside
the pupil. However, these other areas are irregular, so it makes
an analysis of distances of each pixel to background of the
image, figure 1. All operations for isolate the iris is made in a
binary image, where has a 1 if this point is considered possible
pupil and 0 if not [10].

Fig. 1. Input Image and Image binarized.
Vol. 2, No. 11, 2011
34 | P a g e
The pixels with greater distance from the background are
considered part of the pupil, determining the pupil area. See
Figure 2.

Fig. 2. Image of distances, the white represents the greatest distance from
pixel to the bottom; the circle is the highest value.
By means of operations of morphology mathematical of
opening are eliminated small elements (noise) (4)[5][11].
( )
Once detected the pupil, algorithms are used to fill the
pupil, through the closure operation (5) [5][12].
( )
Once obtained and recorded the diameter of the pupil of
each image from the video, the system allows see these results
graphically as shown figures 3 to 11. In the y-axis are the mm
by the pupil diameter, on the axis "X" are represents all the
images from the video, it are get 20 pictures for each second of
video.

Fig. 3. Graphic bars

Fig. 4. Figure continues
In graph with bars and the continuous plot of figures 3 and
4 are shown the diameter of the pupil respectively found in
each of the images [13].
In the figure 5, it was a combination of the graph with bars
and graph continuous.

Fig. 5. Graph combination
In figure 6, red lines represent the average obtained after
twenty images, that is to say, after each second video. In the
graph of the Figure 7 can see how the diameter of the pupil
changes in each second of video.

Fig. 6. Graph with average

Fig. 7. Graph average.
Vol. 2, No. 11, 2011
35 | P a g e
In the graphs of figures 8 and 9, it is possible to recognize
the results found after Interpol, blue bars and black line
segments now represent the images that initially gave the zero.

Fig. 8. Graph Interpolation bars

Fig. 9. Graph Continuous Interpolation
The graph in figure 10 gives the diameter obtained in each
second of video, after the interpolation. In the graph of the
Figure 11 is possible to show the image corresponds to the bar
where is positioned and press the mouse computer.

Fig. 10. Graph Average Interpolation

Fig. 11. Graphical analyzed image
III. RESULTS
The following chart shows the average diameter of the
pupil for each second of video of one of the subjects of test
with predominant emotion joy:

Vol. 2, No. 11, 2011
36 | P a g e

Fig. 12. Figure-1-subject-3
After analyzing the video that corresponds to the graph of
figure 12, observe that when the subject makes laughter
increases the amplitude.

Fig. 13. Images of heights.
Figure 13 contains images that correspond to the high
points of the graph bars in figure 12. The graphs show
interesting results when values of zero diameter. These images
are usually where the person closed the eye or the pupil; the
eye is in the edge, or simply an image that does not contain an
eye, figure 14 shows some examples.

Figure 14 Images zero.
The software can detect the joy and sadness through
graphs, however, is detectable (but it is very difficult) other
emotional states, what requires more extensive analysis to
determine the mixed affective states.
The analysis allows get median, variance, standard
deviation and the range where we find most of the data [14].
IV. CONCLUSIONS
Working with emotions is not easy, the fact feel recorded
or observed does not allow feelings flow easily, although this
is able to detect changes in the iris of individuals, some
researchers have even found not only changes in the eye with
the affective state, also with the degree of cognition [15] [16],
but has not determined a range of dilatation in relation to
affective state. The feeling where became more evident the
change was the joy. It is noticeable from the moment you take
the video, when someone laughs the pupil diameter increases
proportionately, but not directly, on a range of 5 to 7 mm, it is
possible to say that there is a dependency statistics. In some
cases difficulties were encountered because many people tend
to close their eyes when laugh, blink or cry, but these are
changes that also the system can detect.
After a number of tests to the same person are located data
that indicates which mood is reflecting this in their eyes; if
stored the data the system can indicate when the subject is sad
or happy, following the same procedure for each person. It is
possible to consider other types of emotions like anger, stress,
fear, etc.
Once the system can indicate the state of mood of a
subject, we will have to see, that so congruent it is the emotion
of the person with who indicates software; which is complex,
because the subject may lie in their emotional state. However
is clear, the mood changes the diameter of the pupil, allowing
differentiate the joy of sadness.
Vol. 2, No. 11, 2011
37 | P a g e
ACKNOWLEDGMENT
Thanks to The IPN (National Polytechnic Institute) who
through of COFAA (Commission for promotion to the
academic support) allows us to conduct the researches.
REFERENCES
[1] Rosalind W. Picard, Alan Wexelblat, Cliffrod I. Nass ; Conference on
Human Factors in Computing Systems ; CHI '02 extended abstracts on
Human factors in computing systems; Minneapolis, Minnesota, USA ;
Pages: 698 - 699 Year of Publication: 2002 ISBN:1-58113- 454-1
[2] Juan D. Velsquez, Pattie Maes; International Conference on
Autonomous Agents Proceedings of the first international conference on
Autonomous agents; Marina del Rey, California, United States; Pages:
518 - 519 Year of Publication: 1997 ISBN:0-89791-877-0
[3] Rosalind W. Picard, Alan Wexelblat, Cliffrod I. Nass ; Conference on
Human Factors in Computing Systems ; CHI '02 extended abstracts on
Human factors in computing systems; Minneapolis, Minnesota, USA ;
Pages: 698 - 699 Year of Publication: 2002 ISBN:1-58113-454-1
[4] Ric Heishman, Zoran Duric, Harry Wechsler; Using Eye Region
Biometrics to Reveal Affective and Cognitive States, George Mason
University, 2004.
[5] Rodney Brooks; Humanoid robots; Communications of the ACM ;
Volume 45 , Issue 3 ; Robots: intelligence, versatility, adaptivity ; Pages:
33 38,2002.
[6] Paul Ekman, Como detectar menterias, Paidos, 2009.
[7] Costantino Grana; Roberto Mezzani, Rita Cucchiara; Enhancing HSV
histograms with achromatic points detection for video retrieval;
Conference On Image And Video Retrieval; Proceedings of the 6
th
ACM
international conference on Image and video retrieval ;Amsterdam, The
Netherlands ;Pages: 302 - 308 ;Year of
Publication: 2007 ; ISBN:978-1-59593-733-9
[8] Arturo de la Escalera; Visin por Computador Fundamentos y Mtodos;
Prentice Hall Madrid (2001)
[9] Aristide Chikando, Jason Kinser; Optimizing Image Segmentation Using
Color Model Mixtures; 34th Applied Imagery and Pattern Recognition
Workshop Cover; Washington, DC; 2005
[10] Hugo Proenca and Lus A. Alexandre; UBIRIS: A Noisy Iris Image
Database; Lecture Notes in Computer Science Volume
3617,Springer;USA; 2005.
[11] M. Harlick Robert, R. Sternberg Stanlye, Xinhua Zhuang; Image
Analysis Using Mathematical Morphology; IEEE Transactions on
pattern Analysis and Machine Intelligence, Vol. Pam-9, No. 4, July
1997.
[12] De Mira, J, Mayer, J; Image feature extraction for application of
biometric identification of iris - a morphological approach; Computer
Graphics and Image Processing,SIBGRAPI 2003; Brasil, 2003.
[13] Prtega Daz S., Pita Fernndez S; Representacin grfica en el Anlisis
de Datos; Unidad de Epidemiologa Clnica y Bioestadstica; 2001
[14] L. Ipia Santiago, I. Durand Ana; Inferencia estadstica y anlisis de
datos; Pearson Prentice Hal; Espaa; 2008
[15] Ric Heishman, Zoran Duric, Harry Wechsler (2004): Using Eye Region
Biometrics to Reveal Affective and Cognitive States, George Mason
University.
[16] Rodolfo Romero Herrera, Sal De la O Torres, Melody Neftali Rivera
Gutirrez, (2005) ROC&C (IEEE) Deteccin de cambios en el dimetro
de la pupila para el reconocimiento de estados emocionales.
AUTHORS PROFILE
Master degree in computer systems and electronic communications

Vol. 2, No. 11, 2011
38 | P a g e
The impact of competitive intelligence on products
and services innovation in organizations

Phathutshedzo Nemutanzhela
Faculty of Information and Communication Technology
Tshwane University of Technology (TUT)
Pretoria, South Africa
Tiko Iyamu
Faculty of Information and Communication Technology
Tshwane University of Technology (TUT)
Pretoria, South Africa

AbstractThis paper discusses the findings of the study that was
aimed at establishing the effect of Competitive Intelligence,
hereafter referred to as CI, on product and service innovations. A
literature study revealed that CI and innovation are widely
studied subjects, but the impact of CI on innovation was not well
documented. CI has been widely acclaimed as a panacea to a lot
of organizational problems. The study aimed at establishing the
impact of Competitive Intelligence (CI) on products and services
Innovation in organisation. A case study was conducted, using an
Information and communication technology (ICT) organisation.
Innovation-decision process was applied in the data analysis. At
the end of the study on the impact of competitive intelligence on
products and services innovation in organisations the following
was achieved. It was better understood that while CI is
overemphasized as revolutionary, customer focused information
systems products and services still remain challenging. It was also
understood that not all organisations that deploy CI produce
more innovative methods. A lack of knowledge- sharing and
limitations within the organisational culture were found to be
important factors for the deployment of competitive intelligence
products and services in the organisations. Conclusions of the
study basing on the findings are presented.
Keywords-Competitive Intelligence (CI); Diffusion of Innovation
(DoI); Innovation; Product; Services.
I. INTRODUCTION
The ability to innovate is at the heart of every business
survival. When a business runs out of innovative capabilities,
then it is bound to exit the fight to attain and keep customers.
[1] Contends with this view that for over 20 years, successful
product innovation has been considered a key requirement for
business success.
For innovation to realize its full value there is need for the
market to buy the final product or service. This necessitates the
innovating firm to gather as much knowledge as it can about
the needs of the customer, or to create such a need in the
market. This kind of creating a need is what [2] called the need
pull model. In the need pull model the search for a high
potential market initiates a search for the inside or outside the
firm knowledge to develop an innovative product that would
meet market a need. This is the usefulness of Competitive
Intelligence to gather the required knowledge to create
opportunity in the market.
Besides the firm creating a need, there are other needs that
are created as a result of responses to the competition. [3]
contends that Competitive Intelligence units can also help a
firm to understand how a rival has developed their own unique
capabilities and asset caches, assess a rivals ability to imitate
their strategy or assist a company to assess how to uniquely
bundle resources to create value for its customers. This way a
firm will reposition itself to respond with the situation in the
competitive environment. This would be a direct impact of CI.
II. LITERATURE REVIEW
The study aimed to develop a framework for understanding
the impact of competitive intelligence on products and services
innovation in organisations. Related studies were reviewed,
with consideration that this is a conference proceedings paper.
According to Cavalcanti [4] The essence of intelligence
begins with environmental scanning activities, also known as
surveillance. The essence of this process is a transformation of
data, information and knowledge into intelligence as a final
product. However intelligence as a final product becomes
useful if and when the final consumers needs have been
satisfactorily met.
Another value of Competitive Intelligence is the ability to
gather consumer opinion. [4] Asserts that Consumer opinion
can offer insight into the benefits of a product or a service as
well as consumption tendencies. It can also help identify
wishes, dreams and future fantasies. These future fantasies are
the heart of the innovative process. As soon as the organisation
can gain insight of the wishes of the consumers, this will direct
the innovations in an attempt to satisfy the consumer needs.
The extent to which the organisations turn this information into
a resource for product or service innovation is the reason of this
study.
In the literature, it is customary make a distinction between
(competitive) intelligence as a product and as a process (e.g.,
[5, 6, and 7]). In treating intelligence as a product, authors refer
to the information or knowledge obtained and used for
strategic purposes. The process view stresses the process by
means of which this information or knowledge is obtained.
Both the above definitions stress the process aspect. The first
definition also highlights intelligence as a product. If
competitive intelligence is seen as a product, it is usually
compared with data, information and knowledge [5, 6, and 7].
CI has two basic functions namely to move competitive
losses to wins and move competitive pushes (no decisions) to
Vol. 2, No. 11, 2011
39 | P a g e
wins [8]. [8] Asserts that To achieve this kind of impact,
Competitive Intelligence must address the core issues of
departments that have influence on the drivers of competitive
outcomes. Specifically, Competitive Intelligence needs to
address mission critical issues for sales, marketing, and product
development. The issue asserted here is the ability of the users
of Competitive Intelligence to transform whatever information
they have gathered into an innovative product or service. The
best way it will address sales and marketing is providing an
innovative product or service in addition to formulating a
formidable strategy. Hence the purpose of the study was to
establish how far CI contributes to the process of product and
service innovation.
A) CI in the process of products and services innovation
According to some studies, innovation is not a single act
but rather a process which begins with an idea or an invention
[9, 10, 11, 12, and 13]. The classification of Innovations into
different types focuses attention on the outcomes of the
innovation process which have been defined as a new product
or service, a new production process technology, a new
structure or administrative system or a new plan or program
pertaining to organisational members [9].
The power of competitive intelligence to contribute to
product and service innovation lies in its ability to enable
technological forecasting, as [14] aptly titled their 1998 paper
Technological forecasting techniques and Competitive
Intelligence: tools for improving the Innovation process. Their
argument is that Technological forecasting can be used as a
tool to help the forecasters in the process of describing, in the
most precise way possible; how a future machine will be; and
its method of operation and its characteristics. In other words,
how, where and what should be innovated. The information
that serves as the intelligence is that information that is in the
right hands.
Information with a touch of expertise makes a difference in
the process of Innovation.
Frates and Seenas [15] argue that an important subset of
competitive intelligence, competitor intelligence is insufficient,
and potentially misleading. If all a company does is track
known competitors within its own industry, it is likely to suffer
from marketing myopia. This implies that a firm must not only
focus on its competitors, but, more importantly, the whole
general business environment. In essence the goal of CI is to
facilitate a whole rounded body of knowledge that an
organisation can draw from in its formulation of the
Innovations it will present to the customer in form of goods and
services.
However [15] do not explicitly show what impact CI would
have on innovating for the new customers that they refer to.
While they advocate targeting new customer segments, there is
a need to enter these new markets with innovative products that
will capture the markets.
III. RESEARCH METHODOLOGY
The study was descriptive because it was aimed at
answering the question what. Descriptive study tries to
discover answers to the questions who, what, when, where, and
sometimes, how. The researcher attempts to describe or define
a subject [16] The aim was to establish what are the impacts
of Competitive Intelligence on product and service innovation
in organization are. According to Cooper and Schindler [16]
qualitative refers to the meaning, the definition or analogy or
model or metaphor charactering something, while quantitative
assumes the meaning and refers to the measure of it. The
research had a qualitative attribute because this allowed the
researcher to gather perceptions and opinion about Competitive
Intelligence. The quantitative view gathered the proportions of
the practitioners of Competitive Intelligence and be able to
generalize their opinion. Since the study collected opinion
about the impact of Competitive Intelligence in the innovation
process, these opinions were recorded as qualitative data
because reference would be made about them in the form of
quotes. These opinions were gathered through semi-structured
interviews.
There were two main, primary and secondary sources used
in the data collection, namely, interviews and document
review, respectively. According to Leedy and Ormrod [17]
data collection may be anything, such as electronic documents
that can assist the researcher to answer a research question.
Interviews were conducted with identified employees at
Amajita. Relevant documents about the case study were
gathered from authentic sources. This includes access to the
organisations strategy.
The semi-structured interview approach was used.
According to Kvale [18], the most useful interview format for
conducting qualitative research is often semi-structured
(sometimes called moderately scheduled). A total number of
11 interviews were conducted. This number was reached
heuristically, i.e., the decision to stop adding respondents was
taken when nothing new was being learnt from the interviews.
The research main question was: What are the impacts of IS
innovation on CI products and services in organisations? There
were subsidiary question to the main question.
IV. THEORY
The Diffusion of innovations (DoI) theory was used in the
analysis of the data. The notion of stages in an innovation-
decision process conceptualized by [19] was employed.
Therefore a model of the Innovation-decision process that
covers five stages: Knowledge, persuasion, decision,
implementation, and confirmation were used to gain
knowledge and understanding of the impact of competitive
intelligence on information systems innovation products and
services in organisations. This was in order to develop a good
information systems innovation framework.
The analysis was done at two different, macro and micro
levels, individual and group, respectively. Each participant in
the case studies was labelled as follows: in Amajita, as
AJ_LA001 to AJ_LA011.
The Innovation decision process characterized as a process
that occurs while individuals participate in a series of actions
related to decisions [20] Knowledge occurs when individuals
are aware of the Innovation and gain understanding of its
functions. Persuasion is when individuals or decision-making
units exhibit favourable or unfavourable behaviour toward the
Vol. 2, No. 11, 2011
40 | P a g e
Innovation. Decision indicates when the individual or unit
decides to adopt or reject the Innovation. Implementation
occurs when the individual or unit decides to use the
Innovation. Confirmation occurs when decision makers
confirm or reject their decision to adopt the Innovation [20].
V. DATA ANALYSIS
Using the Innovation-Decision Process [19], the analysis of
the data is presented as follows:
A. Knowledge
Knowledge was considered as key and a very important
factor in the organisation. If the organisation is to remain
competitive over time, the managers have to generally see a
need to know more about new products and services,
technology, and about current and potential competitors. One
of the reasons that were attributed to the deployment of
competitive intelligence (CI) is because the market keeps
changing. Also information and communication technology
keep changing and rapidly. Almost on daily basis, there are a
new IS innovation products and services. According to one of
the interviewees AJ_L002 (p8:99-100), the constant and rapid
changing requires awareness by the employees and the
employers, if the organisation is to have competitive advantage.
AJ_LA004 (p20:277-278) also expressed that, the market trend
of IS innovation products and service always changes. The
way things are done changes every day because we are in the
technology world. This challenge us; managers need to
research further and make sure employees are always
knowledgeable about products and services we offer so that we
can remain competitive in the market.
Without the knowledge of the products and services the
organisation wont be able to deploy CI. According to one of
the of the interviewees AJ_LA001 (p4:37-38), we send our
employees to boat camps, when our partner has initiated a new
product and service so that they can be educated of the
challenges faced during the implementation and also get to
know the benefits that the innovation bring into the
organisation.
According to one of the interviewees, AJ_L003 (p32: 475-
480), as an organisation, we face a challenge when employees
are not aware of the products and services that are offered,
because both the organisation and employees are impacted
negatively. The costs of our business get affected and
employees growth is also affected. The employee further
expressed that, Employees, who are not familiar with the
environment they operate on, can be too reluctant and fearful to
take on task. For fear of being exposed, which could lead to
threat to their job, also some employees agrees or volunteer to
take on task, which they lack the knowledge of, this results to
poor quality of service delivery to the clients. This sometimes
results in frustrations on the part of the informed employees
and ultimately resignations. From the organisations point of
view, these employees who lack knowledge were considered
first when faced with retrenchments or dismissals.
B. Persuasion
Successful organisation consists of result-oriented persons.
The capacity to achieve results consistently is tied to develop
the ability to plan. Execute flawlessly, and follow-up
consistently. Effective staffers require improving their personal
effectiveness. It was necessary for each employee to convince
themselves of the products and services. Thereafter, they could
confidently implement and support the products and services
on behalf of the organisation. According to one of the
employees: I think as employees, we must first persuade
ourselves on improving our personal effectiveness towards the
deployment of CI before we can even question our managers
on the measure they are putting in place to persuade
us.(AJ_LA0010 p39:595-596). The employees further
expressed that, even if employer put performance measure in
place to persuade us, the performance will still be poor because
we are not self-effective.
According to one of managers AJ_LA004 (p21:291-292),
we have realized that if an employee is not ready to adopt the
changes brought to them, their performance get affected as they
dont become too involved in the deployment of that products
and services, they even refrained themselves. For organisation
to persuade employees to adopt the deployment of CI
information systems products and service innovation is a
challenge. This is mainly because each employee has his or her
own view of how they see products and services. As such,
many of the interviewees differ in their views and belief. The
views and opinions were informed by individuals and groups
roles in the deployment of CI in the organisation.
According to one of interviewees AJ_LA0011(p42:650-
651), Our organisation get the innovation from a global partner
and it then gets implemented without finding out how we feel
about the innovation, all that the organisation does is to send us
to training so that we can be equipped on the deployment and
benefits of products and services. Therefore employees were
not given room for negotiation. Whatever that gets decided by
the management, the employees have to comply with it. Not
having a room for negotiations was a risk to the organisation, in
terms of losing employees commitment.
AJ_LA009 (p37:558-559) asserted that, Our skills are not
measured, yet our managers give incentives for performance. It
doesnt motivate many employees for them to put all your
effort on a work that you know everyone will be rewarded the
same regardless of the effort, but if the organisation had a
measure in place, I believe all employees were going to strive
for the best deployment. Performance management contract
was required in order to develop a process for managing
individuals and teams to achieve high levels of organisational
performance.
C. Decision
According to AJ_LA001 (p1:9-10), We operate on a
centralised environment, and the innovation to be deployed
comes from our international partner who has already gone
through the whole deployment. He further expressed that our
organisation has to analyse the information of IS products and
service to make sure it aligns with the strategy and culture of
our business, before we can decide to adopt the innovation.
Therefore many of the decisions made were based on
objectives and strategy of the organisation. Analysis is vital in
helping decision-makers understand how phenomena in the
Vol. 2, No. 11, 2011
41 | P a g e
broader environment relate to their companys mission,
objectives, and strategy.
AJ_LA001 (p4:38) mentioned that, We recognized the fact
that organisations may lack the ability or means to respond or
adapt to changes in the environment. This might be as a result
of centralisation, which limits them innovative skills. Also the
complexity and rapidity by which changes occurred had
significant impact on the organisation's competitive challenges.
As a result, an organisation's ability to receive or acquire
information from the environment becomes an essential, yet
not sufficient ingredient in the strategy formulation process.
The partnership with an international company means the
organisation was forced to adopt and deploy competitive
intelligence in the exact way as the partners. This limits the
organisation ability to decide on specific strategy. Decision-
making to them, it just mostly means whether they decide to
implement the product and services or not.
The operation of the organisation was centralised. Among
other things, this was intended to ensure that there was
uniformity of data in the organisation. AJ_LA007 (p32:480-
481) stated that, The business that we operate on is a
centralised operation where all information is stored and
retrieved from one central place called the cloud so everyone
get in the organisation has access to this information which
then makes it easy analysis the data and make decision.
D. Implementation
The organisation preferred to use the CI as the only source
of inspiration and tend to be followers rather than leaders or
innovators within specific markets. According to one of the
employees AJ_LA007 (p30:439-400), Our organisation doesnt
come up with innovation, but rather will follow the footsteps of
our partners, which limit us from being creative and think
strategically. AJ_LA006 (p28:417-418) stated that, It looks
easy and simple to implement products and services which are
already done somewhere, but as the organisation we face
challenges of aligning the innovation with our own culture and
strategy.
The implementation of CI products and services is
influenced by many factors. The factors include: the negative
attitudes of managers, the organisations corporate culture not
being conducive for CI. In some cases, we the managers tend to
think that we know it all. As a result, we fail to listen to advices
from our colleagues, and believe whatever we decide on has to
be implemented exactly that way, the time we realize its
wrong. The damage has been done and all the money used for
implementation has been wasted (AJ_LA003 p11:79-80).
In the organisation, the implementation of CI was a
challenge. This challenge was attributed to lack of strengths
and weakness in context of skills, lack of resources such as
infrastructure, and limited cost. According an interviewee, Like
another competitive intelligence implementation, our
organisation sometimes face challenges of not having enough
resources, sometimes we try to cut implementation cost and
end up missing the important steps of implementation
(AJ_LA001 p3: 22-23).
AJ_LA002 (p10:92-93) Emphasizes on implementation
being the most important stage on this organisation. The
employees view implementation as the most important stage on
competitive intelligence deployment. From strategy to the final
deployment of CI, the organisation adopt a package from USA
which already in use. Which then leaves them with the ability
to only implement the product and service in the South African
environment after adoption? The value that competitive
intelligence add during the implementation process its cutting
cost and time spent on it. Amajita achieve this by using CI
tools that helps them better their services and product delivery.
E. Confirmation
The confirmation of CI products and services in the
organisation was done on case by case. Competitive
Intelligence implementation provides multiple tangible and
intangible benefits to the organisation. However, the realisation
of these benefits depends on how CI was implemented.
According to one of employees: the acceptance of our products
and services depend on how well we implement them, some of
them get accepted well, and we see this through customers that
organisation acquire after deployment (AJ_LA005 p23:314-
315).
The most critical step towards the successful of a CI
solution is an understanding of its users. One of the
interviewees AJ_L009 (p36:546-547), questioned that if
feasibility study was done accurately and all customer needs
were understood before implementation, the organisation had
faced less challenges to whether the customer will accept this
new innovation.
After the implementation of CI on IS products and services
innovation is done, the whole team awaits to see how the
innovation will be accepted or confirmed as a success. The
organisation also go through the own confirmation stage. We
test our innovation and make sure everything is working
according to the project plan (AJ_LA005 p15:359-360). Even
though we have assurance that the innovation works, we still
need a confirmation from our local customer to see if it works
also for them (AJ_LA007 (p32:477- 478).
VI. FINDINGS
From the analysis of the data, eight factors were found to be
critical and influence the deployment of competitive
intelligence products and services in the organisation. The
factors include cultural values, knowledge sharing,
environmental scanning, education and awareness, output,
performance contract, centralisation and negotiation. These
factors are discussed as follows:
a) Cultural Values
The arena of cultural values differs from one environment
to the other. Within cultural settings, people share values about
what is acceptable or not. In this sense, it is very rare for two
organisations which are operating on two different continents
to share the same cultural values. Within such context as
Amajita find itself, this limits the creative skills of those who
their organisation is not dominant. Therefore their culture and
values are compromised to suit the deployment of CI products
and services from another environment. But if this was not the
case, the employees were going to contribute and explore in the
meantime practicing and following their organisation cultural
values. When cultural and values are compromised sometimes
Vol. 2, No. 11, 2011
42 | P a g e
customer lack confidence on their organisational ability to
deliver.
b) Knowledge sharing
Some of information, skills, knowledge and expertise are
being shared in the organisation aimed to improve and sustain
CI products and services. However, there are certain factors
which the organisation hasnt addressed. This factors impacts
on employees willingness to share information. Such factors
are performance contract, organisation perform on a centralized
environment, and fear of losing their jobs if they share
knowledge and skills they have. Their environment limits and
forces the employees to restrain from sharing knowledge, as
they lack measure to use for persuasion. Employees who have
information that is worthy to the organisation, and refrain from
sharing knowledge because of fear that they might lose their
jobs. They also fear that their junior could use it to compete
with them, and fear of losing power that they hold. Not being in
position to innovate, was also another reason why employees
dont share information that can help change the scope of the
business of the organisation. They wait to be told what to do.
Though, organisation has realized that knowledge constitutes a
valuable intangible asset for creating and sustaining
competitive advantages. Knowledge sharing activities were
generally supported by knowledge management systems as
adopted by the organisation. However, technology constitutes
only one of the many factors that affect the sharing of
knowledge in organisations. Others includedorganisational
culture, trust, and incentives. The sharing of knowledge
constitutes a major challenge in the field of knowledge
management because some employees tend to resist sharing
their knowledge with the rest of the organisation. And maybe if
they have measure put in place this can persuade them to share
information. However, trust goes with security, when
employees dont feel secure at their organisation, they wont
have trust. Personal interest will always supersede the interest
of the organisation. As that continue to happens, no one will be
will to share any of expertise that they have.
c) Environmental scanning
The organisation depends on their international partner for
innovation, limiting them to exploit of CI tools. The
organisation uses forms of workshops to check whether their
clients could adopt the innovation which was presented to them
by the international partner. Therefore the relationship they
have built with their external customer plays a bigger role
during environmental scanning. When there is new innovation,
they act as quickly as possible. They should be able to realize
the gap immediately. So that they can either adopt or reject the
innovation. During that time they can utilize their customer as
the tool to acquire information, instead of hosting events that
can cost money and consume time. They have to use them as
tool for finding of information about the environment and to
check if the innovation can fit in. This enables the organisation
to understand the external forces of change so that they may
develop effective responses which secure or improve their
position in the future. Environmental scanning had influence on
the implementation of competitive intelligence products and
services. As a result of the influence, some competitive
intelligence products and services suffer rejection from
potential customer. This is because the organisation didnt take
time to scan the environment, however just implement.
d) Education and Awareness
Education and Awareness plays a significant role in
encouraging and enhancing people's participation in the
activities which were aimed and considered essential for
achieving sustainable competitive intelligence products and
services. The organisation seems to have less emphasis on
educating their staff about the information systems products
and services, particularly, the implementation process of the
deployment of CI products and services. It seems as they just
assume that all employees are aware of the deployment
process. Employees who are not aware of the organisation they
operate to, they are more of an expense, than an asset to the
company. Because they wont even be able to transfer or
acquire skills from each other as they dont have an idea of
whats happening around their organisation. However, if the
organisation can spend time educating and making employees
aware of the CI deployment, they can cut cost and reduce time
spent doing feasibility studies on the process of adopting and
implementing innovation from outside partner.
e) Output
In order to improve the effectiveness and efficiency of CI
products and services, the organisation must consist of result
oriented persons. The capacity to achieve results consistently
was tied to development of the ability to plan, execute
flawlessly, and follow-up consistently. Unfortunately, many
employees in the lacked the ability to plan as they only wait to
be told what to do, instruction from the international partner.
This inability to innovate affects their personal effectiveness.
They can only execute innovation that has already been
executed by the international partner. So they were not needed
to be creative and plan positively. In majority of the cases,
answers were provided to them. However, an organisation is
required to improve their employees personal effectiveness.
Each personnel need to have a good understanding and
motivating strategies competitive intelligence products and
services in the organisation. Because an employee who fails do
discipline themselves affects the performance of the company.
f) Performance
Organisation need to have contract that can sets out the
terms of performance between the employees and the
employer. Because the organisation doesnt have performance
contract in place, employees seem to lack motivation, and this
had impact on CI products and services. They dont trust
colleagues with certain information, skills, knowledge and
expertise they have. This also led to some employees not
sharing information with some of their colleagues. Some of
them even resigned because they feel that they are not secure.
The performance contract builds relationship between
employees and the employer. Because the rights and
obligations of all parties involved in the performance were
covered.
g) Centralisation
The activities of the organisation regarding planning
decision-making they were concentrated within a particular
group. The centralisation approach had impact on innovation of
Vol. 2, No. 11, 2011
43 | P a g e
CI products and services in the organisation. When everything
is centralised in the organisation, the organisation lacks skills to
innovate. This means employee get to be told what to do.
Employees dont have a view on the CI deployment but only to
implement. The organisation force employees to reserve their
input, because everything runs on a centralized space.
Information and innovation comes from centre point, in most
cases, from the international partner.
h) Negotiation
Employees dont have bargaining power. What employer
decides on it goes. The nature of the business (centralised
organisation) limit both employer and employees an
opportunity to negotiate or come up with different views that
could benefit the organisation. Even though the organisation is
dominated by the global market, the culture and values they use
are of the organisation. This put the employees on an
uncomfortable situation where they are deploying a global
innovation, yet the culture and values are not the same and still
cant negotiate. If employees were given a chance to negotiate
they could have had better way to deploy CI products and
services, to suit the culture and values of the South African
environment.
The above findings were interpreted and presented in the
following section.
Interpretation of the Findings
The findings from the analysis are interpreted to be the
factors which influence the innovation of competitive
intelligence (CI) within the information systems environment.
This is depicted in Figure 1 below. The discussions that follow
explain the manifestation from these factors.

Figure 1. Factors Influencing Information Systems Innovation
a) Improvement is a change in the thought process for
doing something, and useful application of new inventions or
discoveries. The organisation doesnt innovate but rather
implement innovation, the environment and nature of the
business that they operate in doesnt give them an opportunity
to come up with innovation, or a discovery of any information
systems products or services. When employees are not aware
of the innovative process (products and services deployment)
they become reluctant and lack innovative and creative
thinking. The employees acts, behaviour and performance
manifest to the organisational performance. This has impact on
skills and knowledge transfer.
b) Dominant the international partner to Amajita was
dominant in the innovation process. They are the one who
comes up with innovation. Its their strategy, plans and
execution that led to the deployment of the competitive
intelligence, making them to be dominant party over Amajita.
While the technology industry dominant technology and
organisation cycles are faster than in many other organisations,
the point is still that dominant organisations have to evolve or
they will fail. Therefore, Global innovator should provide the
full array of competencies needed by successful CI
practitioners. Therefore Amajita leaders and managers must
understand the significance of these sources in South African
market, rather than agree to the naive concept that they get
from their international partners as the market cannot be
exactly the same.
c) Distribution - An organisation or set of organisations
(go-betweens) involved in the process of making a product or
service available for use or consumption by a consumer or
business user. By sharing knowledge about products and
services, employees gain more than they lose. Sharing
knowledge is a synergistic process you get more out than you
put in.
d) Collaboration - is a process where two or more
people or organisations work together to realise shared goals,
(this is more than the intersection of common goals seen in co-
operative ventures, but a deep, collective, determination to
reach an identical objective). If you try to work alone you are
likely to fail you need not only the input from other people
but their support and buy-in. Being open with them; sharing
with them, helps you achieve your objectives.
e) Reliance employees might trust transferring skills to
others, and know that the other employees and the organisation
at large would benefit from it. They should also trust that the
information wont be used against them tomorrow. Which
means organisation has to assure them of security and that they
job wont be at stake if they are willing to share information
and expertise they have.
f) Involvement is the state of a user in relation to a
situation after that user has participated in the situation. It is,
approximately, an internal analogue to aconflict of interest.
Even with the low level of knowledge sharing that goes on
today if you do not make your knowledge productive than
someone else with that same knowledge will. You can almost
guarantee that whatever bright idea you have someone else
somewhere in the organisation will be thinking along the same
lines.
VII. CONCLUSION
It was established that the role of Competitive Intelligence
on product and service innovation was to inform strategic
management, reflect customer needs, and inform rivals about
the competitors and help firms locate themselves on the
competitive scale.
Vol. 2, No. 11, 2011
44 | P a g e
It was established Competitive Intelligence is
overemphasized as revolutionary yet customers still remain
unsatisfied by the services and products because most of
Competitive Intelligence remains in talk, but not in execution.
This was reflected by the numbers of respondents who said that
Competitive Intelligence was not useful because what was
collected was not used.
The findings and the analysis of the study indicate that
further research relating to competitive intelligence can be
conducted. Some of the suggestions are:
Organizational Culture it would be in the interest of
academic organisations to investigate and gain a better
understanding of how organisational culture impacts on
Competitive Intelligence products and services in
organisations.
Underpinning theories it would be a significant
contribution to apply different theories such as Organisational
information processing theory (OIPT) and Technology-
Organisation-Environment Framework (TOT framework) in a
similar study. Organisation need to be knowledgeable about the
Competitive intelligence products and services they deploy so
that they can make better decisions to innovate. Organisational
information processing theory is applied to Organisations that
need quality information to cope with environmental
uncertainty and improve their decision making [21].
REFERENCES
[1] R. L. Chapman, C. E. Omara , S. Ronchi S., A., M. Corso,. Continuous
product innovation: a comparison of key elements across different
contingency sets. Measuring Business Excellence. 2001 Pp 16-23.
Available from www.emerald-library.com/ft [Accessed 10/08/2010]
[2] A. R. Burgelman, L. R. SAYLES. Transforming invention into
innovation: the conceptualization stage in Burgelman A. R., christensen
C. M., wheelwright S. C. 2004. Strategic management of technology and
innovation 4Ed. McGraw Hill-Irwin Boston.
[3] S. Hughes. Competitive Intelligence as Competitive Advantage: The
Theoretical Link between Competitive Intelligence, Strategy and Firm
Performance. Journal of Competitive Intelligence and Management . Vol
3. No 3. Available at www.scip.org [Accessed 20/08/2010]
[4] P. E. Cavalcanti. The Relationship between Business Intelligence and
Business Success Journal of Competitive Intelligence and Management
.Vol 3. No 1 available at www.scip.org [Accessed 20/08/2010]
[5] L.M. Fuld. The New Competitor Intelligence. Chichester. UK: Wiley,
1995.
[6] L. Kahaner. Competitive Intelligence. New York: Touchstone, 1997.
[7] E. A. Philips, and D. Vriens. Business Intelligence. Deventer: Kluwer,
1999.
[8] R. B. Jeppsen. An introduction to impact-based Competitive
Intelligence. 2005 www.scip.org [Accessed 16/ 05/10]
[9] F. Damanpour. Organizational innovation: a meta-analysis of effects of
determinants and moderators. Acad. Mgmt. J.1991. pp.34:555.90.
[10] D. Dougherty, and E. Bowman. The effects of organizational
downsizing on product innovation. California Management
Review,1995. pp.37: 28-44.
[11] F. Drucker and F. Peter. The discipline of innovation, Harvard Business
Review,80,(8). 2002. pp.77-83.
[12] Tidd, Joseph, Pavitt, Keith (contributor), Tidd, Joe, R. Bessant.
ManagingInnovation : Integrating Technological, Market and
Organizational Change, 2nd edition, John Wiley & Son Ltd, UK. 2001.
[13] A. H. Van de ven, Polley, E. Douglas, Garud, Raghu, Venkataraman,
Sankaran.The Innovation Journey, Oxford University Press, USA.2000.
[14] . D. Lemos, and A. C. Porto. Technological forecasting techniques
and Competitive Intelligence: tools for improving the Innovation process
Industrial Management & Data Systems. 1998. pp7:330337.
[15] J. Frates and S. Seena. Using Business Intelligence to Discover New
Market Opportunities, Journal of Competitive Intelligence and
Management, vol. 3, no. 2, 2005. available at www.scip.org [Accessed
20/03/2010]
[16] D. R. Cooper and P. S. Schindler. Business research methods 8th ed.
McGraw Hill, Boston. 2003.
[17] P. D. Leedy and J. E. Ormrod. Practical research: planning and design.
8th ed.: Pearson. MAREE K., 2007. First steps in research. Pretoria:
Van Schaik.2009.
[18] S. Kvale. Interviews: An introduction to qualitative research London:
Sage.1996.
[19] E. M. Rogers. Diffusion of Innovations. 4th ed. New York: Free Press.
1995.
[20] E. M. Rogers. Diffusion of innovations (5th ed.). New York: Free Press.
2003.
[21] G. Premkumar, K. Ramamurthy and C. S. Saunders. Information
processing view oforganisations: An exploratory examination of fit in
the context of interorganisational relationships. Journal of Management
Information Systems, 22(1), 2005.pp.257-294).
[22] Phathutshedzo Nemutanzhela and Tiko Iyamu; A Framework for
Enhancing the Information Systems Innovation: Using Competitive
Intelligence.
AUTHORS PROFILE
Phathutshedzo Nemutanzhela has a Masters degree (MTech) Business
Information Systems from Tshwane University of Technology. She has a
Baccalaureus Technologies (BTech): Information Technology
(Informatics). Her principle research interest is Competitive Intelligence,
Information Systems and innovation. Theoretically, she focuses on
Diffussion of Innovation (DoI) Theory.
Tiko Iyamu is a Professor of Information Systems at the Tshwane University
of Technology, Pretoria. His research interests include Mobile
Computing, Enterprise Architecture and Information Technology
Strategy. Theoretically, he focuses on Actor Network Theory (ANT) and
Structuration Theory (ST). Iyamu is author of numerous peer-reviewed
journal and conference proceedings articles.

Vol. 2, No. 11, 2011
45 | P a g e
Arabic Sign Language (ArSL) Recognition System
Using HMM

Aliaa A. A.Youssif , Amal Elsayed Aboutabl, Heba Hamdy Ali
Department of Computer Science, Faculty of Computers and Information
Helwan University
Cairo, Egypt

AbstractHand gestures enabling deaf people to communication
during their daily lives rather than by speaking. A sign language
is a language which, instead of using sound, uses visually
transmitted gesture signs which simultaneously combine hand
shapes, orientation and movement of the hands, arms, lip-
patterns, body movements and facial expressions to express the
speaker's thoughts. Recognizing and documenting Arabic sign
language has only been paid attention to recently. There have
been few attempts to develop recognition systems to allow deaf
people to interact with the rest of society. This paper introduces
an automatic Arabic sign language (ArSL) recognition system
based on the Hidden Markov Models (HMMs). A large set of
samples has been used to recognize 20 isolated words from the
Standard Arabic sign language. The proposed system is signer-
independent. Experiments are conducted using real ArSL videos
taken for deaf people in different clothes and with different skin
colors. Our system achieves an overall recognition rate reaching
up to 82.22%.
Keywords-Hand Gesture; Hand Tracking; Arabic Sign Language
(ArSL); HMM; Hand Features; Hand Contours.
I. INTRODUCTION
Singing has always been part of human communications
[1]. For millennia, deaf people have created and used signs
among themselves. These signs were the only form of
communication available for many deaf people.Within the
variety of cultures of deaf people all over the world, signing
evolved to form complete and sophisticated languages. These
languages have been learned and elaborated by succeeding
generations of deaf children.
Normally, there is no problem when two deaf persons
communicate using their common sign language. The problem
arises when a deaf person wants to communicate with a non-
deaf person. Usually both will be dissatisfaction in a very short
time.
In this section we focus our discussion of the efforts made
by researchers on sign language gesture recognition in general
and on Arabic sign language (ArSL) in particular. Sign
language recognition systems can be further classified into
signer-dependent and signer-independent. Also one may
classify. Sign language recognition systems are either glove-
based which relies on electromechanical devices for data
collection, or none glove-based if free hands are used. The
learning and recognition methods used in previous studies to
recognize sign language include neural networks and hidden
Markov models (HMMs).
Cyber gloves have been widely used in most of previous
Works on sign language recognition including [1, 2, 3]. Kudos
[4] reported a system using power gloves to recognize a set of
95 isolated Australian sign languages with 80% accuracy.
Grobel and Assan [5] used HMM to recognize isolated signs
with 91.3% accuracy out of a 262-sign vocabulary. They
extracted 2D features from video recordings of signers wearing
colored gloves. Colored gloves were used in [6] where HMM
was employed to recognize 52 signs of German sign language
with a single color video camera as input. In a similar work [7]
an accuracy of 80.8% was reached in the corpus of 12 different
signs and 10 subunits using the K-means clustering algorithm
to get the subunits for continuous sign language recognition.
Liang and Ouhyoung [8] employed the time-varying parameter
threshold of hand posture to determine end-points in a stream
of gesture input for continuous Taiwanese sign language
recgnation with the average recognition rate of 80.4% over 250
signs. In their system HMM was employed, and data gloves
were taken as input devices.
The use of cyber gloves or other means of input devices
conflicts with recognizing gestures in a natural context and is
very difficult to run in real time. Therefore, recently
researchers presented several sign recognition systems based
on computer vision techniques [9, 10, 11]. Starner et al [12]
used a view-based approach for continuous American
continuous sign language recognition. They used a single
camera to extract two-dimensional features as the input of
HMM. The accuracy of 92% or 98% was obtained when the
camera was mounted on the desk or in a users cap in
recognizing the sentences with 40 different signs. However, the
user must wear two colored gloves (a yellow glove for the right
hand and an orange glove for the left) and sits in a chair before
the camera. Vogler and Metaxas [14] used computer vision
methods and interchangeably AT a flock of birds for 3D data
extraction of 53 signs for American Sign Language. They,
respectively, built context-dependent HMM and modeled
transient movement to alleviate the effects of movement
epenthesis. Experiments over 64 phonemes extracted from 53
signs showed that modeling the movement epenthesis has
better performance than context-dependent HMM. The system
provided overall accuracy of 95.83% is reported.

Vol. 2, No. 11, 2011
46 | P a g e
Furthermore, most of the above systems are signer
dependent systems. A more convenient and efficient system is
the one that allows deaf users to perform gestures naturally
with no prior knowledge about the user. To the best of our
knowledge there are a few published works on signer-
independent systems. Vamplew and Adams [13] proposed a
signer-independent system to recognize a set of 52 signs. The
system used a modular architecture consisting of multiple
feature-recognition neural networks and the nearest neighbor
classifier to recognize isolated signs. They reported a
recognition rate of 85% in the test set. Again the signer must
wear cyber gloves while performing gestures. Another attempt
is made by Fang et al [14] in which they used the SOFM/HMM
model to recognize signer-independent CSL over the 4368
samples from 7 signers with 208 isolated signs.
ArSLs are still in their development stages. A glove-based
and singer-dependant Arabic sign recognition system has been
developed by M. Mohandes and S. I. Quadri, M. Deriche [15].
They used a data set of 15 samples for each of the 300 signs
which were carried out by a signer wearing a pair of colored
gloves (orange and yellow) achieving recognition accuracy
about 88.73%. Jarrah and Halawani [3] developed a system for
ArSL alphabet recognition using a collection of Adaptive
Neuro-Fuzzy Inference Systems, a form of supervised learning.
They used images of bare hands instead of colored gloves to
permit the user to interact with the system conveniently. The
used feature set comprised lengths of vectors that were selected
to span the fingertips region and training was accomplished by
the use of a hybrid learning algorithm achieving recognition
accuracy of 93.55%. Likewise, Assaleh and Rousan [1]
extended the work in [3] by using Polynomial classifiers
extracted superior results on the same dataset. Their work
required the participants to wear gloves with colored tips while
performing the gestures to simplify the image segmentation
stage. They extracted features including the relative position
and orientation of the fingertips with respect to the wrist and to
each other. The resulting system achieved 93.41% recognition
accuracy.
More recently, recognition of video-based isolated Arabic
sign language gestures is reported by in [16] and [17]. In [16],
the dataset is based on 23 gestures performed by 3 signers. The
data collection phase did not impose any restrictions on
clothing or background. Forward or bi-directional prediction
error of the input video sign was accumulated and threshold
into a single image. The still image is then transformed into the
frequency domain. The feature vector that represents the
gesture is then based on the frequency coefficients. Simple
classification techniques such as KNN, linear and Bayesian
were used. This work was extended in [17] where a block-
based motion estimation technique was used to find motion
vectors between successive images. Such vectors are then
rearranged into intensity images and transformed into the
frequency domain.
This paper is organized as follows. Section II describes the
Arabic sign language database used in the work. Section III
describes the hand features. Hand tracking and recognition
phases of the proposed system are elaborated in Section IV.
Section V presents the modeling of the Arabic Sign Langue
using Hidden Markov Models. The experiments and results are
discussed in Section VI. Finally, conclusion and future work
are presented in Section VIII.
II. ARABIC SIGN LANGUAGE (ARSL) DATABASE
Because there has been no serious attention to Arabic sign
language recognition, there are no common databases available
for researchers in this field. Therefore, we had to build our own
database with reasonable size. As depicted in Fig. 1, in the
video capturing stage, a single digital camera was used to
acquire the gestures from signers in a video format. At this
stage, the video is saved in the AVI format in order to be
analyzed in later stages. When it comes to recognition, the
video is streamed directly to the recognition engine.
The database in this work is collected in collaboration with
ASDAA Association for Serving the Hearing Impaired
(ASDAA) [18]. The videos are captured from the deaf
community who volunteered to perform the signs to generate
samples for our study. The database consists of a 20-word
lexicon given in Table 1.
No restrictions are imposed on the signer or word length.
The words pertain to common situations in which handicapped
people might find themselves in. The database itself consists of
45 repetitions of each of the 20 words performed by different
signers, 20 of which are used for training and 18 for testing. No
restriction is imposed on clothing, background, age or sex of
the signer. Moreover, signers are gloves-free and with different
signers with different skin colors. It was totally free hands. The
deaf sign word is captured using a digital video camera. The
frame rate was set to 25 frames per second with a spatial
resolution of 640x480.
III. HAND FEATURES
The image features, together with information about their
relative orientation, position and scale, are used for defining
understated but discriminating view-based object model [19].
We represent the hand by a model consisting of (i) the palm
as a coarse scale blob, (ii) the five fingers as ridges at finer

Figure 1. Sample videos of our Arabic sign language (ArSL) database.
Vol. 2, No. 11, 2011
47 | P a g e
scales and (iii) finger tips as even finer scale blobs as in Fig. 2.
We then define different states of the hand model, depending
on the number of open fingers.

To model translations, rotations and scaling transformations
of the hand, a feature vector is defined to describe different
hand features including the global position y), size and
orientation and its discrete state.
IV. HAND TRACKING AND RECOGNITION PHASES OF
PROPOSED SYSTEM
In this paper, a system for recognizing Arabic sign
language gestures is presented. There are three main phases for
hand detection and tracking; skin detection, edge detection and
hand fingertips tracking.
A. Skin Detection
Each video contains a collection of frames representing a
gesture. At first, each video is pre-processed by applying a
video segmentation technique that captures frames with a
frame rate of 25Hz. Then, the RGB captured frames are
converted into HSV image because it is more related to human
color perception [20]. These color spaces separates three
components: the hue (H), the saturation (S) and the brightness
(V). Essentially, HSV-type color spaces are deformations of
the RGB color cube .They can be mapped from the RGB space
via a nonlinear transformation. One of advantages of these
color spaces in skin detection is that they allow users to specify
the boundary of the skin color class in terms of the hue and
saturation. As V gives the brightness information, they are
often dropped to reduce illumination dependency of skin color.
Given an image, each pixel in the image is classified as a
skin or non - skin using color information. The histogram is
normalized and if the height of the bin corresponding to H and
S values of a pixel exceeds a threshold called skin threshold
(obtained empirically), this pixel is considered a skin pixel.
Otherwise, the pixel is considered a non-skin pixel. A general
image and its skin detected image can be seen such that white
pixels represent the hand gesture and black pixels represent the
background or any object behind the skin as shown in Fig. 3
(a). Finally, smoothing is applied to each frame using a
MEDIAN filter to remove noise and shadow.
B. Canny Edge Detection
The Canny algorithm uses an optimal edge detector based
on a set of criteria which include finding the most edges by
minimizing the error rate, marke edges as closely as possible to
the actual edges to maximize localization and marke edges
only once when a single edge exists for minimal response [21].
C. Hand Contours and Fingertips Tracking
Hand tracking is the process of locating a moving hand (or
both hands) over time using a camera. For each frame extract,
the contours of all the detected skin regions in binary image
using connected component analysis are detected. Tests are
performed to detect whether the input contour is convex or not.
The contour must be simple, i.e. without self-intersections. The
signers head is considered to be the biggest detected region
and the moving hand as the second biggest region. Features
considered include the position of the head, coordinates of the
center of the hand region and direction angle of the hand
region. Other features that represent the shape of the hand are
also considered and are extracted from changes of image
intensities called image motion:
( ) ( ( ) (( )) (1)
Thus, the next frame recorded at time can be obtained
by moving every point in the current frame, recorded at time ,
by suitable amount. The amount of motion ( ) is called
displacement of the point at ( ).
The displacement vector is a function of the image position
, and variations in it are often noticeable even within the
small tracking window. We try to find interesting points with
big eigenvalues in an image to be added to the feature vector.
Word
Number
Arabic Meaning English Meaning
1 Alexandria
2 Drink
3 Eat
4 Me
5 You
6 English
7 Association
8 Deaf
9 Arabic
10 French
11 Book
12 Happy
13 Talkative
14 Confused
15 Uncomfortable
16 School
17 Responsible
18 Egypt
19 Home
20 Sleep

Figure 2. Feature-based hand models in different states. The circles and
ellipses correspond to blob and ridge features. When aligning models to
images, the features are translated, rotated and scaled according to the
feature vector
TABLE I. RRECOGNITION RATES OF DIFFERENT HMM MODELS
WITHDIFFERENT FEATURE VECTOR LENGTHS
Vol. 2, No. 11, 2011
48 | P a g e
These interesting points (corners) are characterized by a large
variety in all directions of the vector . By analyzing the
eigenvalues of the image pixels, this characterization can be
expressed in the following way: we should have two "large"
eigenvalues for an interesting point. Based on the magnitudes
of the eigenvalues, the following inferences can be made based
on this argument:
If 10 and 2 0 (the two interesting points"
eigenvalues for an inter, st point) then this pixel (x,y) has no
features of interest. In this case we reject the corners with the
minimal eigenvalue less than quality Level
If 1 and 2 have large positive values, then a corner is
found. The Shi-Tomasi [22, 23] corner detector directly
computes min (1, 2) because under certain assumptions, the
corners are more stable for tracking.
Finally, it ensures that all the corners found are distanced
enough on from another by considering the corners (the
strongest corners are considered first) and checking that the
distance between the newly considered feature and the features
considered earlier is larger than the minimum distance. So, the
function removes the features than are too close to the stronger
feature.
An example of interesting points found that represent a
motion is shown in Fig. 3( b).In this work, the implementation
of Hand tracking method is carried out using OpenCV (Open
Source Computer Vision); a library of programming functions
for real time computer vision. [24].
V. MODELLING OFARSL USING HMM
HMMs (Hidden Markov Models) have been prominently
and successfully used in sign languages. HMM is a
probabilistic model representing a given process with a set of
states (not directly observed) and transition probabilities
between the states. Such a model has been used in a number of
applications including the recognition of the Sign Language
recognition [25, 26].
Let each sign be represented by a sequence of gestures or
observations O, defined as:
O = o1, o2,.........., ot (2)
Where o
t
is the feature vector observed at time t. The sign
recognition problem can then be regarded as that of computing:
arg maxi {P(wi|O)} (3)
Where w
i
is the ith vocabulary word. This probability is
not computable directly but using Bayes Rule [27]:
H(e
i
,O) =( H(O,e
i
)H(e
i
))/ H(O) (4)
Thus, for a given set of prior probabilities P (wi), the most
likely sign depends only on the likelihood P (O|wi). Given the
dimensionality of the observation sequence O, the direct
estimation of the joint conditional probability P (o1, o2 |wi)
from examples of sign is not possible. However, if a parametric
model of word production such as a Markov model is.
As shown in Fig 4, each gesture for a sign is modeled as a
single HMM with N observations per gesture (o1, o2, ot). In
HMM based sign recognition, it is assumed that the sequence
of observed feature vectors corresponding to each gesture is
generated by a Markov model as shown in Fig 4. A Markov
model is a finite state machine which changes state once every
time unit and each time t that a state j is entered, a feature
vector ot is generated from the probability density bj(ot).
Furthermore, the transition from state i to state j is also
probabilistic and is governed by the discrete probability aij .
Fig 4 shows an example of this process where the six state
model moves through the state sequence X = 1; 2; 2; 3; 4; 4; 5;
6 in order to generate the sequence o1 to o6. It is to be noted
that the entry and exit states of a HMM are non-emitting in the
Hidden Markov Model Toolkit which is used in this work [28].
This facilitates the construction of composite models as
explained in more detail later.
The joint probability that O is generated by the model M
moving through the state sequence X is calculated simply as
the product of the transition probabilities and the output
probabilities. So is for the state sequence X in Fig 4.
A. Training Phase
Training in the context of our work means learning or
generating an HMM given a sequence of observations. For
each training sequence XT11 , ...,XTn1 , ...,XTN1 of a gesture
of class k with N sequences, the image features are prepared
and then extracted .

(a) (b)
Figure 5. the result of computing blob features and ridge features from
an image of a hand. (a) Result image after skin detection (b) circles and
ellipses corresponding to the significant blob and features extracted from
an image of a hand; it describes how the selected image features capture
the essential structure of a hand.

Figure 3. State HMM model for gesture "Alexandria"
Vol. 2, No. 11, 2011
49 | P a g e
These extracted images are used as feature vectors in the
Viterbi [15] training to train a hidden Markov model k for
each gesture as shown in Fig. 5.
B. Recognition Phase
This phase involves finding the probability of an observed
sequence given an HMM and finding the sequence of hidden
states that most probably generated an observed sequence. The
feature extraction of the test sequences is identical to the
training process. Then for each test pattern the hidden Markov
model which best describes the current observation sequence is
searched as shown in Fig. 6.

The implementation of the HMM for our ArSL system has
been carried out using the HTK toolkit [29]. HTK is a portable
toolkit for building and manipulating hidden Markov models.
HTK is primarily used for speech recognition research
although it has been used for numerous other research areas
including speech synthesis, character recognition and DNA
sequencing. HTK is in use at hundreds of sites worldwide.
VI. EXPERIMENTS AND RESULTS
The training method, described earlier, has been
implemented by creating one HMM model per class (gesture),
resulting in a total of 20 models.
Several experiments have been conducted to evaluate our
ArSL recognition system. All experiments are performed on
the same training data collected in prior. Depending on the
same database, the system attempts to recognize all samples for
every word where the total number of samples considered here
is 360. We use 6 HMM models which have different number of
states and different number of Gaussian mixtures per state. For
each experiment, the 6 models are used for different feature
vector lengths (5, 8 and 9) achieving recognition rates of
78.61%, 82.22% and 80.27 % respectively as shown in table 2.

The set of experiments shown in table 2 has been conducted
for each of the 20 Arabic signs in our database. The best result
(recognition rate) obtained for each sign along with the
associated best model is shown in table 3.
It is noticeable that some signs result in particularly low
recognition rates. The gesture eat, for example, has a
recognition rate of 55.56% and is mostly classified as drink.
This is due to the fact that the location, movement and
orientation of the dominant hand are very similar in both
gestures. Therefore, the observation (feature) vectors, o1,
o2,., ot produced from the feature hand tracking phase
are most likely very close to each other. Thus, the system will
get confused between these two signs and provide relatively
higher error rate for these particular gestures. A similar
situation occurs with the sign Me which has a recognition
rate of 66.67% is mostly classified as {You}.
Our ArSL proposed recognition system is based on 8
features per frame which is considered better than the
previously published results in the field of ArSL while Jarrah
and Halawani [3] use of 30 elements as a length of feature
vector per video frame. M. Rousan and K. Assaleh [15] use
feature vector of 50 elements.
We compare our work with that done in [28]. In spite of the
fact that they use different feature extraction methods, setup,
and database, both systems are based on the same classifier
(HMM). As shown in Table 4, our proposed system (referred to
as ArSL-Using HMM in Table4) system performs much better
than the DCT coefficient-based system (referred to as the ArSL
DCT coefficient-based system in Table 4).
ArSL DCT coefficient-based system uses 50 elements
length of feature vector and Recognition rate 90.6%.It is
expected that the increase in a feature vector size be
accompanied by a corresponding increase in recognition rates.
This is due to the fact that each DCT coefficient is uncorrelated
with other coefficients and hence no redundant information is
presented in increasing coefficients.

HMM model No of features
5 elements 8 elements 9 elements
[3states/10mixture] 47.78% 58.89% 56.67%
[6states/2mixture] 41.94% 45.83% 55.56%
[4states/2mixture] 34.44% 30% 30%
[6states/6mixture] 61.94% 66.11% 68.61%
[6states/4 mixture] 56.94% 56.94% 43.05%
[6states/10mixture] 75.55% 71.94% 74.16%
Overall Recognition Rate 78.61% 82.22% 80.27%

Figure 5. Traning phase

Figure 6. Recognition Phase
TABLE II. RECOGNATION RATES OF DIFFERENT HMM MODELS WITH
NUMBER OF STATES AND MIXTURE
Vol. 2, No. 11, 2011
50 | P a g e

VII. CONCLUSION AND FUTURE WORK
We have demonstrated that our Arabic sign language
recognition system is effective considering the nature of the
videos used and the number of features considered. Our system
is signer-independent. The database that we built consists of
videos taken for deaf people using their normal life Arabic sign
language. The signers are gloves-free, with varying clothes and
skin colors. Importantly, only 8 features have been considered
which is less than the number of features used previously by
other researchers. The overall recognition rate is 82.22%,
which is reasonably high considering the number of features
used.
In the future, we aim to achieve higher recognition rates
with a larger data set. A psycholinguistic study on the structure
of Arabic sign language might be needed to choose the
appropriate HMM model (the perfect number of states) for
each gesture. We will explore and test our training models to
build a continuous sentence recognition system using a sub-
gesture word based recognition system. Such a system will
help the deaf community to interact and integrate with the rest
of the society.
REFERENCES
[1] K. Assaleh and M. Al-Rousan, Recognition of Arabic Sign Language
Alphabet Using Polynomial Classifiers EURASIP Journal on Applied
Signal Processing, 2005 (13): 2136-2146,2005.
[2] W. Gao, J.Y. Ma, J.Q. Wu, C.L. Wang, Sign language recognition
based on HMM/ ANN/DP, International Journal of Pattern Recognition
Artificial Intelligent 14 (5) (2000) 587602.
[3] O. Al-Jarrah, A. Halawani, Recognition of gestures in Arabic sign
language using neuro-fuzzy systems, Artificial Intelligence 2 (133) 117
138, 2001.
[4] M.W. Kadous, Machine recognition of Auslan signs using
PowerGloves: towards large-lexicon recognition of sign language, in:
Proceedings of the Workshop on the Integration of Gestures in
Language and Speech, pp. 165174, 1996.
[5] K. Grobel, M. Assan, Isolated sign language recognition using hidden
Markov models, in: Proceedings of the International Conference on
System, Man and Cybernetics, pp. 162167, 1997.
[6] H. Hienz, B. Bauer, K.F. Krais, HMM-based continuous sign language
recognition using stochastic grammar, in: Proceedings ofGW99, LNAI
1739, pp. 185196, 1999.
[7] B. Bauer, K.F. Kraiss, Towards an automatic sign language recognition
system using subunits, in: Proceedings of the International Gesture
Workshop, pp.6475, 2001.
[8] R.H. Liang, M. Ouhyoung, A real-time continuous gesture recognition
system for sign language, in: Proceedings of the Third International
Conference on Automatic Face and Gesture Recognition, pp. 558565,
1998.
[9] V.I. Pavlovic, R. Sharma, T.S. Huang, Visual interpretation of hand
gestures for humancomputer interaction: a review, IEEE Transactions
on Pattern Analysis and Machine Intelligence 19 (7) 677695, 1997.
[10] J.J. Triesch, C. Malsburg, A system for person-independent hand
posture recognition against complex backgrounds, IEEE Transactions on
Pattern Analysis and Machine Intelligence 23 (12) 14491453, 2001.
[11] P. Vamplew, A. Adams, Recognition of sign language gestures using
neural networks,Australian Journal of Intelligent Information Processing
Systems 5 (2) 94102, 1998.
[12] T. Starner, J. Weaver, A. Pentland, Real-time American sign language
recognition using desk and wearable computer-based video, IEEE
Transactions on Pattern Analysis and Machine Intelligence 20 (12)
13711375, 1998.
[13] Y. Wu, T.S. Huang, Vision-based gesture recognition: a review, in:
Proceedings of the International Gesture Workshop, pp. 103115, 1999.
[14] G.L. Fang, W. Gao, J.Y. Ma, Signer-independent sign language
recognition based on SOFM/HMM, in: Proceedings of the IEEE ICCV
Workshop Recognition, Analysis and Tracking of Faces and Gestures in
Real-Time Systems, pp. 9095, 2001.
[15] M. Mohandes, S. I. Quadri, M. Deriche "Arabic Sign Language
Recognition an ImageBased Approach" 21st International Conference
on Advanced Information Networking and Applications Workshops
(AINAW'07) 0-7695-2847-3/07 $20.00 2007.
[16] T. Shanableh, K. Assaleh and M. Al-Rousan, Spatio- Temporal
Feature-Extraction Techniques for Isolated Gesture Recognition in
Arabic Sign Language IEEE Trans. On Systems, Man and Cybernetics
Part B, 37 (3): 641-650, 2007.
[17] T. Shanableh and K. Assaleh, Telescopic Vector Composition and
polar accumulated motion residuals for feature extraction in Arabic Sign
Language recognition EURASIP Journal on Image and Video
Processing, vol. 2007, Article ID 87929, 10 pages, 2007.
[18] ASDAA Association For Serving The Hearing Impaired (ASDAA)
[19] Lars Bretzner1, 2 Ivan Laptev1, Tony Lindeberg "Hand Gesture
Recognition using Multi-Scale Colour Features, Hierarchical Models
and Particle Filtering Proceedings of the Fifth IEEE International
,Arabic
Meaning
English
Meaning
Best Result

HMM [Number of
States / Number of
Gaussian mixtures]
Alexandria
77.78%
[6states/10 mixture]
Drink 88.89% [3states/10 mixture]
Eat 55.56% [3states/10 mixture]
Me
83.33%
[6states/10 mixture]
You 66.67% [6states/10 mixture]
English 88.89% [6states/10 mixture]
Association 66.67% [6states/6 mixture]
Deaf 88.89% [6states/10 mixture]
Arabic 88.89% [6states/6 mixture]
French 94.44% [6states/4 mixture]
Book 55.56% [6states/10 mixture]
Happy 100% [6states/6 mixture]
Talkative 72.22% [6states/10 mixture]
Confused 100% [6states/10 mixture]
Uncomfortable 88.89% [6states/6 mixture]
School 77.78% [6states/10 mixture]
Responsible 100% [6states/10 mixture]
Egypt 77.78% [6states/6 mixture]
Home 72.22% [6states/6 mixture]
Sleep 100% [3states/10 mixture]
Instruments
used
Feature vector
length
Recognition
Rate
ArSL
DCT
coefficient-
based
system
[28]
None: free
hands
50 elements of
DCT
coefficients per
frame
90.6%
ArSL-
Using
HMM
None: free
hands
considering the
head position
8 features per
frame
82.22%
TABLE IV. COMPARISON WITH SIMILAR SIGNER-INDEPENDENT, HMM-
BASED SYSTEMS.
TABLE III. RECOGNITION RATES WITH DIFFERENT MODELS FOR EVERY
WORD WITH BEST MODEL
Vol. 2, No. 11, 2011
51 | P a g e
Conference on Automatic Face and Gesture Recognition (FGR.02) 0-
7695-1602-5/02 $17.00 IEEE 2002
[20] A. Albiol, L. Torres, and E. J. Delp. Optimum color spaces for skin
detection. In proceedings of the 2001 international conference on
image processing, volume 1, vol. 1, pp. 122-124 ,2001.
[21] Ravikiran J, Kavi Mahesh, Suhas Mahishi, Dheeraj R, Sudheender S,
Nitin V Pujari "Finger Detection for Sign Language Recognition"
Proceedings of the International MultiConference of Engineers and
Computer Scientists 2009 Vol I IMECS 2009, March 18 - 20, Hong
Kong, 2009.
[22] J. Shi and C. Tomasi, "Good Features to Track,". 9th IEEE Conference
on Computer Vision and Pattern Recognition. Springer, June 1994.
[23] C. Tomasi and T. Kanade, "Detection and Tracking of Point Features".
Pattern Recognition 37: 165168, 2004.
[24] Open Source Computer website. [Online]. Available Vision
http://opencv.willowgarage.com/
[25] Prof. Dr.-Ing. H. Ney, Appearance-Based Gesture Recognition
Diplomarbeit im Fach Informatik Rheinisch-Westfalische Technische
Hochschule Aachen Lehrstuhl fur Informatik VI, 2005
[26] Khaled Assaleh, Tamer Shanableh, Mustafa Fanaswala, Harish Bajaj,
and Farnaz Amin, "Vision-based system for Continuous Arabic Sign
Language Recognition in user dependent mode Proceeding of the 5th
International Symposium on Mechatronics and its Applications
(ISMA08), Amman, Jordan, May 27-29, 2008.
[27] Forney GD, The Viterbi algorithm". Proceedings of the IEEE 61 (3):
268278. doi:10.1109/PROC.1973.9030, 1973.
[28] M. AL-Rousan, K. Assaleh , A. Talaa. Video-based signer-
independent Arabic sign language recognition using hidden Markov
models Applied Soft Computing 9 990999, 2009.
[29] The HTK website. [Online]. Available http://htk.eng.cam.ac.uk
AUTHORS PROFILE
Aliaa A. A.Youssif, professor of computer science. in
Faculty of Computers and Information, Helwan University.
Cairo, Egypt. She received her B.Sc and MSc. degree in
telecommunications and electronics engineering from
Helwan University. Dr. A. Youssif received the PhD degree
in computer science from Helwan University in 2000. She
was a visiting professor at George Washington University
(Washington DC, USA) in 2005. She was also a visiting
professor at Cardiff University in UK (2008). Her fields of interest include
pattern recognition, AI researches, and medical imaging. She published more
than 40 papers in different fields.

Amal Elsayed Aboutabl is currently an Assistant Professor
at the Computer Science Department, Faculty of Computers
and Information, Helwan University, Cairo, Egypt. She
received her B.Sc. in Computer Science from the American
University in Cairo and both of her M.Sc. and Ph.D. in
Computer Science from Cairo University. She worked for
IBM and ICL in Egypt for seven years. She was also a
Fulbright Scholar at the Department of Computer Science, University of
Virginia, USA. Her current research interests include parallel computing,
performance evaluation and image processing.

Heba Hamdy Ali received her BSc. degree in computer
science from Helwan University. She is currently a masters
degree student under supervision of prof. Aliaa A. A.Youssif
and Dr Amal Elsayed Aboutabl. Her areas of interests
include image processing, pattern recognition, HMM and
artificial intelligence.

Vol. 2, No. 11, 2011
52 | P a g e
Modularity Index Metrics for Java-Based Open
Source Software Projects

Andi Wahju Rahardjo Emanuel
Informatics Bachelor Program,
Faculty of Information Technology,
Maranatha Christian University,
Bandung, Indonesia
Retantyo Wardoyo, Jazi Eko Istiyanto,
Khabib Mustofa
Dept. of Computer Science and Electronics,
Universitas Gadjah Mada,
Yogyakarta, Indonesia

Abstract Open Source Software (OSS) Projects are gaining
popularity these days, and they become alternatives in building
software system. Despite many failures in these projects, there
are some success stories with one of the identified success factors
is modularity. This paper presents the first quantitative software
metrics to measure modularity level of Java-based OSS Projects
called Modularity Index. This software metrics is formulated by
analyzing modularity traits such as size, complexity, cohesion,
and coupling of 59 Java-based OSS Projects from sourceforge.net
using SONAR tool. These OSS Projects are selected since they
have been downloaded more than 100K times and believed to
have the required modularity trait to be successful. The software
metrics related to modularity in class, package and system level
of these projects are extracted and analyzed. The similarities
found are then analyzed to determine the class quality, package
quality, and then combined with system architecture measure to
formulate the Modularity Index. The case study of measuring
Modularity Index during the evolution of JFreeChart project has
shown that this software metrics is able to identify strengths and
potential problems of the project.
Keywords-Open source software projects; modularity; Java;
sourceforge; software metrics; system architecture.
I. INTRODUCTION
Open Source Software (OSS) Projects are gaining
popularity these days. They were once only considered as an
experimental way of academics and researchers to share the
programming experiences, now they become the mainstream
software development methodology comparable to those of
commercial and proprietary software projects. This movement
was initially started by Richard Stallman [33] and Eric
Raymond [31]. Some success stories of OSS Projects include
Linux Operating System, Apache Web Server, Mozilla Web
Browser, LibreOffice, etc. The success of these projects is
attributed to many key success factors such as the fact that the
developer is the actual user [10], and sound and modular
architecture [20][17][11], the existence of communities that
support the system development [9], etc. From all these success
factors, modularity of the software system is one of the
important factors to be examined further in this paper.
Even though there are some proofs of the success of OSS
Projects, some facts that many more similar projects are
unsuccessful or failed also unavoidable exist [16]. There are
some characteristics of OSS Projects that have been identified
contributing to such unfruitful result such as no formal means
i.e. no project planning [4], poor coding styles of project
initiators [13] and poor architectural design [12]. We believe
that some new approaches with respect to modularity to
counter such problems in OSS Projects are needed. Until now,
modularity has been identified as a key success factor of OSS
projects, but how to apply modularity, especially from early
phase of the project is not yet understood.
This paper presents the formulation of Modularity Index
which is the first quantitative software metrics to measure the
modularity level in OSS Projects.
This paper is organized as follows: section 2 describes the
recent studies in OSS Projects, modularity in OSS Projects and
Software Metrics. Section 3 describes the data source of OSS
Projects for analysis. Section 4 shows the step by step
Modularity Index formulation starts from class level, package
level, and system level. The case study of 33 out of 52
versions of JFreeChart projects is shown in section 5. Finally,
section 6 describes the conclusion of the paper and future
studies of the research.
II. RECENT STUDIES
A. OSS Projects
Many web portals have been developed as an incubator for
OSS Project's developers to develop and host their projects.
These portals are equipped with many development tools and
statistics to assist the project initiator or administrator in
improving their projects and other interested contributors to
join the projects. Some of the popular portals are
Sourceforge.net, freshmeat.net, launchpad.net, and Google
Code.
The OSS Projects themselves have several distinct
characteristics not found in commercial / proprietary software
development [10][26], which are:
- The source code of the application is freely available
for everybody to download, improve and modify [31].
- People who contribute to the development of the OSS
projects are usually forming a group called
communities. The recruitment process if this groups
are completely voluntary [9]. This communities is an
example of true merit-based system of hierarchy [11]
Vol. 2, No. 11, 2011
53 | P a g e
- The development methods of the projects are lacking of
formal methodology found in commercially developed
software applications [4]. The two most important
activities are fixing bugs and adding features [3].
There are already many studies relating to OSS Projects
that are classified into three main categories. The first category
is the study of large and successful OSS Projects to find their
success characteristics such as Debian [32], FreeBSD [12],
Apache [27], Open BSD [22], and many more. The second
category is the study to find similarities in several OSS Projects
such as Apache dan Mozilla [26], 15 OSS Projects [35], and 2
OSS Projects [6]. The last category is the study on the process
aspects in OSS Projects such as Requirement Engineering [30],
code fault [22], Design Pattern [18], reliability model [37],
phase of development [34], and work practice in OSS projects
[10].
Current studies about OSS Projects mostly focus on the
already successful and large projects that have already
established hierarchy and system, while most of the failed and
unsuccessful OSS Projects are usually small or medium sized
projects [16]. The application of these hierarchy and system in
already established projects into small to medium sized projects
may not be suitable. In our initial research, we have conducted
analysis on more than 130K OSS Projects to find their success
factors [15].
B. Modularity in OSS Projects
Modularization involves breaking up of an software system
into smaller, more independent elements known as module
[23]. Booch has defined modularity as the property of a system
whose modules are cohesive and loosely-coupled [24]. Fenton
stated that modularity is the internal quality attribute of the
software system [24]. It is also known that modularity is
directly related to software architecture, since modularity is
separation of a software system in independent and
collaborative modules that can be organized in software
architecture [29]. Modular software has several advantages
such as maintainability, manageability, and comprehensibility
[28]. Moreover, modularity has been identified as one of the
key success factors in OSS Projects [20][17][11].
There are five attributes closely related to modularity in
software system which are coupling / dependency, complexity,
cohesion, and information hiding [21][7]. To have an ideal
modular software system, the system should have the following
attributes:
- Small size in each module (package) and many
modules in the system [36]: each module / package
should only responsible for simple feature, and the
more complex features should be composed of many of
these simple features. The possible software metrics to
measure size are NCLOC (non-commenting lines of
code), Lines, or Statements.
- Low coupling / dependency [5]: minimization or
standardization of coupling / dependency e.g. through
standard format i.e. published APIs [2], elimination of
semantic dependencies, etc.
- Low complexity: hierarchy of modules that prefers
flatter than taller dependency [28][2].
- High cohesion [21]: high integrity of the internal
structure of software modules which is usually stated as
either high cohesion or low cohesion.
- Open for extension and close to modification [5]:
capability of the existing module to be extended to
create a more complex module. And avoid changing
already debugged code. The creation of new modules
should be encourage using available extension and not
modifying the already tested module.
Even though modularity is already identified as the key
success factor in OSS Projects, the justification for it in large
and succesful OSS Projects is purely qualitative. The software
metrics attributing to the modularity properties are all separated
and not yet integrated into a single measure. This paper will
present a single measure called Modularity Index that
quantitatively determines the modularity level of OSS Projects.
C. Software Metrics
Software metrics are defined as certain values which are
expressed in some units attributed to software application [25].
The software metrics are useful in indicate the current state of
the software and enable to compare and predict the current
achievement of software applications [25]. There are several
known software metrics based on its categories [25]:
- Size-related software metrics: NCLOC, Memory
footprint, Number of classes / headers, Number of
methods, Number of attributes, Size of compiled code,
etc.
- Quality-related software metrics: Cyclomatic
complexity, Number of states, Number of bugs in
LOC, Coupling metrics, Inheritance metrics, etc.
- Process-related software metrics: failed builds, defect
per hour, requirement changes, programming time,
number of patches after release, etc.
There are currently more than 200 metrics with many
different purposes [25], and one of the study by the authors are
the statistical analysis of software metrics affecting modularity
in OSS Projects [14].
III. DATA SOURCE OF OSS PROJECTS
The data source of the OSS Projects for the experiment is
from the sourceforge.net portal since it is the largest OSS
Portal.
A. Assumptions and Considerations
There are several consideration and assumption in selecting
which OSS Projects to be analyzed, which are:
- The OSS projects are build using Java programming
language, and a single package in the project resembles
a module in modular software system. The addition
of package in the software is intended as the addition of
new feature in the system.
Vol. 2, No. 11, 2011
54 | P a g e
- The project's size is limited to small-to-medium-sized
OSS Projects. The limitation of the size (NCLOC) of
OSS Projects being evaluated are 170K. The concept of
modularity is a lot easier to comprehend in object-
oriented programming language (i.e. C++, Java, etc.)
compared to procedural programming (i.e. C, Fortran,
etc.), since the concept of module, coupling, cohesion,
etc. are more straightforward. Java-based OSS Projects
are selected since they are among the mostly popular
object oriented programming for developing Open
Source Software [16].
- The Projects should already be downloaded more than
100,000 times. This high number of downloads may
indicate the success of the projects, which in turn
may imply modularity traits that already identified as
the success factor of OSS Project [20][17][11].
- The source code of the OSS Project is syntax error-free
and compile-able. The SONAR tool requires that the
source code should be compiled first using compile
tool such as maven, or ant. Many of the OSS Projects
provides separate binary and source code and it is
difficult to create binary directly from the source code
due to several reasons such as compile error, build tool
configuration error, syntax error, etc.
B. Selected OSS Projects
Table 1. shows the list of OSS Projects as a subject for this
research. The initial OSS Projects to be evaluated are 209
projects, but only 59 which are suitable to be evaluated using
SONAR due to the assumptions and considerations stated in
section III.A. There are total 1885 modules / packages being
measured from these 59 OSS Projects.
TABLE I. LIST OF 59 SELECTED OSS PROJECTS
No Project Name No Project Name
1 FreeMind 31 Jin client for chess servers
2 jEdit 32 SAX: Simple API for XML
3 TV-Browser - A free EPG 33 jKiwi
4 JFreeChart 34 Data Crow
5 JasperReports - Java Reporting 35 Wicket
6
OpenProj - Project
Management
36 Cewolf - Chart TagLib Project
7 HyperSQL Database Engine 37 DrawSWF
8 yura.net 38
c3p0:JDBC DataSources
/ Resource Pools
9 JabRef 39 JavaGroups
10 FreeCol 40
OmegaT - multiplatform CAT
tool
11
jTDS - SQL Server and Sybase
JDBC driver
41 FreeGuide TV Guide
12 Torrent Episode Downloader 42 Eteria IRC Client
13 FindBugs 43 MeD's Movie Manager
14 PMD 44 subsonic
15 JGraph Diagram Component 45 kXML
No Project Name No Project Name
16 ANts P2P 46 Jaxe
17 Paros 47 The JUMP Pilot Project
18
ProGuard Java Optimizer and
Obfuscator
48
Aglet Software Development
Kit
19 TripleA 49 Antenna
20 JSch 50 CBViewer
21 Jajuk 51 Sunflow Rendering System
22 FreeTTS 52 Thingamablog
23
A Java library for
reading/writing Excel
53 BORG Calendar
24 checkstyle 54
Directory Synchronize Pro
(DirSync Pro)
25 httpunit 55 Java Treeview
26 JMSN 56 Java Network Browser
27 PDFBox 57 Red Piranha
28 JBidwatcher 58 Cobertura
29 JTidy 59 Jake2
30 Jena - -
C. Steps
In order to be able to analyze these OSS Projects, there are
some steps being performed, which are:
- Compile the source code using available build tool
(Ant or Maven2).
- Execute maven2 script to start analyze the OSS
Projects using SONAR tool.
- Creating custom portal to perform the required
analysis.
- Analyze and find the correlation and similarities of all
the projects such as using scatter graph, least square fit,
histogram, etc.
IV. MODULARITY INDEX FORMULATION
The formulation of modularity index will start from the
class level, then move up to the package level, and finally
concluded in the system level.
A. Class Level Modularity
There are four software metrics that determine the level of
modularity in class level, which are:
- Size Metrics which consists of: NCLOC, Lines, and
Statements. NCLOC is the number of non-commenting
lines of code. The selection of NCLOC will also
represent the other size metrics [14].
- Cohesion: LCOM4 or Lack of Cohesion Method
version 4, this version is better for object oriented
programming such as Java as proposed by Hitz and
Montazeri [19] which is the improvement of LCOM1
Chidamber and Kemerer [8].
- Complexity: McCabe's Cyclomatic Complexity [22] is
one example of complexity metrics that widely used.
Vol. 2, No. 11, 2011
55 | P a g e
Our previous paper have shown that the size metrics
and complexity metrics are highly related so this
metrics may be ignored [14].
- Functions: the number of functions / methods in the
class. This may indicates the complexity
1) NCLOC: Figure 1 shows the histogram of the class vs.
NCLOC of the all OSS Projects being evaluated. The value of
NCLOC peaked at 50 with the histogram before the peak
resembles linear straight line and after the peak resembles
inverse polynomial line. The value of approximation of both
lines are shown in the Fig.1.

Figure 1. Histogram of Classes vs. NCLOC
If LOC
Q
is defined as the normalized value of the quality of
NCLOC, so the formula of LOC
Q
are:
LOC
Q
= 0.0125 x NCLOC + 0.375 for NCLOC 50 (1)
LOC
Q
= (NCLOC 50)
-2.046
for NCLOC > 50 (2)
Where:
LOC
Q
= NCLOC Quality Value
NCLOC = NCLOC Value
Note: the value of constant in formula (1) is adjusted from
0.371 into 0.375 to achieve the maximum value of 1 at NCLOC
= 50.
2) Number of Functions: Figure 2 shows the histogram of
classes vs. functions of all OSS Projects being evaluated. The
peak value is 4.83 (rounded up into 5). Similar to class vs.
NCLOC, the values before the peak resembles a straight line
and after the peak resembles an inverse polynomial line with
the approximation of both lines shown in the Fig.2.
F
Q
is defined as the normalized value of function's quality,
it can be formulated as follows:
F
Q
= 0.172 x F + 0.171 for F 5 (3)
F
Q
= (F 4.83)
-2.739
for F > 5 (4)

Figure 2. Histogram of Classes vs. Functions
Where:
F
Q
= Function Quality Value
F = Number of Function
3) Cohesion: Cohesion is determined by the value of
LCOM4. The ideal value is 1 which means that the class is
highly cohesive. Higher value of LCOM4 indicates the degree
of needed separation of classes into smaller classes.
LCOM4 1 (5)
Where:
LCOM4 = Class Cohesion Value
4) Class Quality Formulation: Integrating all above
measures into a single normalized value, the formulation of
class quality or C
Q
are:
(6)
Where:
C
Q
= Class Quality Value
LOC
Q
= NCLOC Quality Value
F
Q
= Function Quality Value
LCOM4 = Class Cohesion Value
B. Package Level Modularity
Package Quality or P
Q
is the quality of individual package.
Since in a single package there are many classes and there is no
similarities found the the optimal number of classes in each
package, so the Package Quality is determined by the average
Class Quality or stated as:
P
Q
= avg(C
Q
) (7)
Where:

Vol. 2, No. 11, 2011
56 | P a g e
P
Q
= Package Quality Value
C
Q
= Class Quality Value
C. System Level Modularity
S
A
is a normalized value (with maximum value of 1) which
determine the value of software architecture. The factors that
influence this value are Package Cohesion (relationship among
classes within package) and Package Coupling (relationship
among classes from different packages). The principle used
here is Maximize Cohesion and Minimize Coupling which
becomes a widely known principle in building a good software
system. The form of formulation is based on presentation titled
Software Architecture Metrics by Ammar et. al [1], with the
difference is that instead of using entropy approaches, this
formulation is using the actual value of dependencies in
determining the value of Package Cohesion and Package
Coupling.
(8)
Where:
Cii = Package Cohesion
Cij = Package Cohesion + Package Coupling
(if i=j is Package Cohesion,
if i j is Package Coupling)
d = number of package
D. Formulation of Modularity Index
Finally, the formulation of Modularity Index is the product
of SA and the sum of all package quality in the software system
as stated in the following formula:
(9)
Where:
M
I
= Modularity Index
S
A
= Software Architecture Value
P
Qi
= Package Quality of Package i
The proposed modularity index is a quality metrics will
have the following properties:
- It has no upper bound: the value of modularity index
increases as the number of module / package increases.
- The value of modularity index, especially the value of
S
A
depends on how the packages are coupled to each
other. The limitation of connection of packages to only
itself (package cohesion) or to only some dedicated
packaged (e.g APIs, proxy, etc.) will improve the value
of S
A
.
V. CASE STUDY: JFREECHART
JFreeChart is a free 100% Java chart library that makes it
easy for developers to display professional quality charts in
their applications (http://www.jfree.org/jfreechart) . This
projects is one of the 59 OSS Projects used for modularity
index formulation. For this case study, this project is chosen
because:
- High S
A
value (more than 0.7 since version 0.9.21)
- Relatively large number of packages (more than 30)
There are 52 versions available from the project's site, but
only 33 are able to be analyzed using SONAR tool and being
measured. The results are show in the following Fig.3.

Figure 3. Average PQ in 33 versions of JFreeChart
Fig. 3 above shows that the average package quality of the
JFreeChart over 33 versions are decreasing consistently. This
indicates the problem in the quality of each classes in each
packages, such as:
- increasing size of NCLOC in each class.
- increasing number of functions in class.
- decreasing number of LCOM4 (Cohesion Metrics) in
class.

Figure 4. SA value in 33 versions of JFreeChart

Vol. 2, No. 11, 2011
57 | P a g e
Fig. 4 above shows that the structure of software
architecture is improving. After consistent decrease in SA
value in early versions of the system, there seems to be
significant effort conducted before the release of version 1.0.0
started from version 0.9.21. The system from version 0.9.21
onward showing high number of SA.
The modularity index itself is shown in Fig. 5. The figure
is showing improvement by the factor of two from early
versions (until version 0.9.20) and late versions (version 1.0.5
onwards). There are significant jump in the value of
modularity index from version 0.9.21 until version 1.0.2
indicating the period of major restructuring of the system
before the release of milestone version 1.0.0.

Figure 5. Modularity Index in 33 versions of JFreeChart
It can be seen from above case study that Modularity Index
and its components (P
Q
and S
A
) are able to point the strength
and potential problems in the development of JFreeChart OSS
Projects. This information may give a valuable insight to the
initiator and developers of the project in improving their
project.
VI. CONCLUSION
Open Source Software (OSS) Projects are now gaining
popularity and becoming one alternatives in developing
software. Despite the the many success story of OSS Projects
such as Apache, Mozilla, etc., the fact the many more of these
projects that are failed needs are alarming. Some studies have
identified that modularity is one of the key success factors of
OSS Projects and authors believe that implementing modularity
approach since early start of the project will increase the
success of the project. This paper presents the first quantitative
measure of modularity for Java-based OSS Projects called
modularity index.
The formulation of modularity index are performed by
analyzing the software metrics attributing to modularity of 59
Java-based OSS Projects from sourceforge.net which have
been downloaded more than 100K times. By analyzing the
similarity of these projects from class level, package level, and
system level, the modularity index are formulated. As the
validation of the software metrics, 33 out of 52 versions of
JFreeChart OSS projects are analyzed using this metrics and
the metrics are able to identify the strength and potential
problems of the project.
Future study relating to this metrics involve further
validation and integration into a framework called modularity
framework in which the measurement of Modularity
Frameworks will generate recommendations for improvement
during OSS projects development. The integration of the
software metrics into a web-based IDE will provide useful tool
for project initiators and developers in improving their OSS
Projects.
ACKNOWLEDGMENT
The authors would like to thank Maranatha Christian
University (http://www.maranatha.edu) that provides funding
for the research, and the Department of Computer Science and
Electronics, Universitas Gadjah Mada (http://mkom.ugm.ac.id)
that provides technical support for the research.
REFERENCES
[1] Ammar H., Shereshevsky M., Mili A., Rabie W., Radetsky N. (2008),
Software architecture metrics, Seminar Presentation, Faculty of
Information Science & Engineering, Management & Science University,
Shah Alam, Malaysia, May 12, 2008. Available:
http://www.docstoc.com/docs/6802629/Software-Architecture-Metrics
[2] Aruna M., M.P. Suguna Devi M.P, Deepa M. (2008), Measuring the
quality of software modularization using coupling-based structural
metrics for an OSS system, Proceeding of the First International
Conference on Emerging Trends in Engineering and Technology 2008
[3] Asundi J. (2007), The need for effort estimation model for open source
software projects, Proceeding of the Open Source Application
Workspace: Fifth Workshop on Open Source Sofware Engineering 2007
[4] Bouktif S., Antoniol G., Merlo E. (2006), A feedback based quality
assessment to support open source software evolution: the GRASS case
study, 22nd IEEE International Conference on Software Maintenance
2006, pp 155 - 165
[5] Cai Y., Huynh S. (2007), An evolution model for software modularity
assessment, Proceeding of the Fifth International Workshop on
Software Qualty 2007 (WoSQ'07).
[6] Capiluppi A., Ramil J.F. (2004), Studying the evolution of open source
systems at different levels of granularity: two case studies, Proceeding
on the 7
th
IEEE International Workshop of Principles of Software
Evolution, 2004, pp 113 - 118.
[7] Capra E., Francalanci C., Merlo F. (2008), An empirical study on the
relationship among software design quality, development effort, and
governance in open source projects, IEEE Transactions on Software
Engineering Vol. 34, No. 6, Nov/Dec 2008, pp 765 782.
[8] Chidamber S.R., Kemerer C.F. (1994), Metrics suite for object
oriented design, IEEE Transaction on Software Engineering, Vol. 20
No. 6 June 1994, pp 476 493.
[9] Christley S., Madey G. (2007), Analysis of activity in the open source
software development community, Proceeding of the 40th IEEE
Annual Hawaii International Conference on System Sciences, 2007, pp
166b.
[10] Crowston K., Wei K., Li Q., Howison J. (2006), Core and periphery in
free / libre and open source software team communications, Proceeding
of the 39th IEEE Hawaii International Conference on System Sciences
2006
[11] DeKoenigsberg G. (2008), How successful open source projects work,
and how and why to introduce students to the open source world, 21st
IEEE Conference on Software Engineering Education and Training,
2008, pp 274 276.
[12] Dinh-Trong T., Bieman J.M. (2004), Open source software
development: a case study of FreeBSD, Proceedings of the 10th IEEE
International Symposium on Software Metrics, 2004, pp 96 - 105.
[13] Ellis H.J.C., Morelli R.A. , Lanerolle T.R., Damon J., Raye J. (2007),
Can humanitarian open-source software development draw new sudents
to CS?, Proceeding of the 38th SIGCSE Technical Symposium on
Computer Science Education 2007, pp 551 555.

Vol. 2, No. 11, 2011
58 | P a g e
[14] Emanuel A.W.R, Wardoyo R., Istiyanto J.E., Mustofa K. (2011),
Statistical analysis on software metrics affecting modularity in open
source software, International Journal of Computer Science and
Information Technology (IJCSIT), Vol. 3, No. 3, June 2011, pp 105 -
118
Success rules of OSS projects using datamining 3-itemset association
rule, International Journal of Computer Science Issue (IJCSI), Vol. 7
Issue 6 Nov. 2010, pp 71 80.
Success factors of OSS projects from sourceforge using datamining
association rule, Proceeding of 2010 International Conference on
Distributed Frameworks for Multimedia Applications (DFmA) 2010, pp
141 - 148
[17] Gurbani V.K., Garvert A., Herbsleb J.D. (2005), A case study of open
source tools and practices in commercial setting, Proceeding of the fifth
Workshop on Open Source Software Engineering 2005, pp 1 - 6.
[18] Hahsler M. (2005), A quantitative study of the adoption of design
patterns by open source software developers, Chapter V of Free / Open
Source Software Development by Stefan Koch, Idea Group Publishing,
ISBN 1-59140-371-5, 2005, pp 103 123.
[19] Hitz M., Montazeri B. (1995), Measuring coupling and cohesion in
object-oriented systems, Proceeding International Symposium on
Applied Corporate Computing, Oct. 25-27 1995, Monterrey, Mexico,
75-76, 197, 78-84
[20] Lawrie T., Gacek C. (2002), Issues of dependability in open source
software development, Software Engineering Notes vol 27 no 3 of
ACM Sigsoft. May 2002. Pp 34 -37
[21] Lee Y., Yang Y., Chang K.H. (2007), Metrics & evolution in open
source software, Proceeding on Seventh International Conference on
Quality Software 2007
[22] Li P.L., Herbsleb J., Shaw M. (2005), Finding predictors of field
defects for open source software systems in commonly available data
sources: a case study of OpenBSD, Proceeding of 11th IEEE
International Software Metrics Symposium, 2005, 32.
[23] McCabe T. (1976), A complexity measure, IEEE Transactions On
Software Engineering, Vol. Se-2, No. 4, December 1976, pp. 308-320.
[24] Melton H., Tempero E. (2007), Toward assessing modularity,
Proceeding of the First International Workshop on Assessment of
Contemporary Modularization Techniques 2007 (ACoM'07)
[25] Meyer B., Oriol M., & Schoeller B. (2009), "Software engineering:
lecture 17-18: estimation techniques and software metrics, Chair of
Software Engineering Website, available:
http://se.inf.ethz.ch/teaching/2008-S/se-0204/slides/15-Estimation-and-
metrics-1-6x.pdf , accessed: 18 January 2009
[26] Mockus A., Fielding R.T., Herbsleb J.(2002), Two case studies of open
source software development: apache and mozilla, ACM Transaction
on Software Engineering and Methodology Vol. II No. 3, Juli 2002, 309
346
[27] Mockus A, Fielding R.T, Herbsleb J.(2000), A case study of open
source software development: the apache server, ACM ICSE, 2000,
263 272
[28] Munelly J., Fritsch S., Clarke S. (2007). An aspect-oriented approach to
the modularisation of context. Proceedings of the Fifth Annual IEEE
International Conference on Pervasive Computing and Communication
(PerCom'07)
[29] Nakagawa E.Y, de Sousa E.P.M., de Britto Murata K. (2008), Software
architecture relevance in open source software evolution: a case study,
Annual IEEE International Computer Software and Application
Conference, 2008, pp 1234 1239.
[30] Paech B, Reuschenbach B (2006), Open source requirements
engineering, Proceeding of 14th IEEE International Requirement
Engineering Conference: 257 - 262
[31] Raymond E.S. (2000), The cathedral and the bazaar, version 3,
Thyrsus Enterprises (http://www.tuxedo.org/~esr/), 2000.
[32] Spaeth S., Stuermer M. (2007), Sampling in open source development:
the case for using the debian GNU/linux dstribution, Proceedings of the
40th IEEE Hawaii International Conference on System Sciences, 2007,
pp 166a.
[33] Stallman R. (1992), Why software should be free, GNU Websites, 24
April 1992, Available:
http://www.gnu.org/philosophy/shouldbefree.html
[34] Stewart K. J., Darcy D.P., Daniel S.L. (2005), Observations on patterns
of development in open source software projects, Proceeding on the
fifth Workshop on Open Source Software Engineering 2005, pp 1 5.
[35] von Krogh G., Spaeth S., Haefliger S. (2005), Knowledge reuse in open
source software: an exploratory study of 15 open source projects,
Proceeding of 38th Hawaii International Conference on System
Sciences, 2005, pp. 198b
[36] Wang Y., Shao J. (2003), Measurement of the cognitive functional
complexity of software, Proceedings of the Second IEEE International
Conference on Cognitive Informatics 2003 (ICCI'03).
[37] Zhou F., Davis J. (2008), A model of bug dynamics for open source
software, The Second IEEE International Conference on Secure
System Integration and Reliability Improvement 2008, pp 185 - 186.
AUTHORS PROFILE

Andi Wahju Rahardjo Emanuel is a Full Time
Lecturer at the Bachelor Informatics Program, Faculty
of Information Technology, Maranatha Christian
University in Bandung, Indonesia. He is graduated as
BSEE in Purdue University, Indiana, USA in 1996 and
MSSE in The University of Melbourne in 2001. He is
currently taking his Doctoral Program at the
Department of Computer Science and Electronics,
Gadjah Mada University in Yogyakarta, Indonesia

Retantyo Wardoyo is an Associate Professor at the
Universitas Gadjah Mada in Yogyakarta, Indonesia.
He is graduated as Bachelor of Mathematics in Gadjah
Mada University, Indonesia . He received his M.Sc in
Computer Science in University of Manchester, UK
and received his PhD in Computation in University of
Manchester Institute of Science and Technology, UK.

Jazi Eko Istiyanto is a Professor and Head of the
He is graduated as Bachelor of Physics in Gadjah
Mada University, Indonesia. He gets his Postgraduate
Diploma (Computer Programming and
Microprocessor), M.Sc (Computer Science) and PhD
(Electronic System Engineering) at University of
Essex, UK.

Khabib Mustofa is an Assistant Professor at the
He is graduated as Bachelor of Computer and Master
of Computer at Gadjah Mada University, Indonesia.
He receives his Dr. Tech in Computer Science at The
Vienna University of Technology, Austria.

Vol. 2, No. 11, 2011
59 | P a g e
Survey of Nearest Neighbor Condensing Techniques

MILOUD-AOUIDATE Amal (corres.Author)
Computer Sciences Department
University of Sciences and Technology Houari
Boumediene, USTHB,
Algiers, ALGERIA
BABA-ALI Ahmed Riadh
Electronic Department
University of Sciences and Technology Houari
Boumediene, USTHB,
Algiers, ALGERIA

AbstractThe nearest neighbor rule identifies the category of an
unknown element according to its known nearest neighbors
categories. This technique is efficient in many fields as event
recognition, text categorization and object recognition. Its prime
advantage is its simplicity, but its main inconvenience is its
computing complexity for large training sets.
This drawback was dealt by the researchers community as the
problem of prototype selection.
Trying to solve this problem several techniques presented as
condensing techniques were proposed. Condensing algorithms try
to determine a significantly reduced set of prototypes keeping the
performance of the 1-NN rule on this set close to the one reached
on the complete training set. In this paper we present a survey of
some condensing KNN techniques which are CNN, RNN, FCNN,
Drop1-5, DEL, IKNN, TRKNN and CBP.
All these techniques can improve the efficiency in computation
time. But these algorithms fail to prove the minimality of their
resulting set. For this, one possibility is to hybridize them with
other algorithms, called modern heuristics or metaheuristics,
which, themselves, can improve the solution. The metaheuristics
that have proven results in the selection of attributes are
principally genetic algorithms and tabu search. We will also shed
light in this paper on some recent techniques focusing on this
template.
Keywords- Nearest neighbor (NN); kNN; Prototype selection;
Condensed NN; Reduced NN; Condensing; Genetic
algorithms; Tabu search.
I. INTRODUCTION
The K-nearest neighbor classification rule (KNN)
proposed by T. M. Cover and P. E. Hart [4], is a powerful
classification method that allows an almost infallible
classification of an unknown prototype through a set of
training prototypes. It is widely used in pattern recognition
[20] [18], text categorization [10] [6], object recognition [8]
and event recognition [23] applications.
An inevitable consequence of large sets of prototypes is
the computational time implied by this research problem. The
databases, used in some areas such as intrusion detection, are
constantly and dynamically updated. This constitutes one of
the main inconveniences of the KNN rule. Another important
inconvenience comes from the fact that the training prototypes
can contain noisy or mislabeled prototypes that may affect the
results and distort them. The scientific community has tackled
these problems and proposed a selection of prototypes which
could modify an initial set of prototypes by reducing its size in
order to improve the classification performance.
II. PROTOTYPE SELECTION
Prototype selection is the process of finding representative
patterns from the data, which can help in reducing these data.
This problem is classified as an NP-hard problem by many
researchers [1] [25], because there is no polynomial algorithm
allowing for the solution. The existing algorithms can just give
acceptable solutions.
Like many other combinatorial problems, the prototype
selection (PS) would require an exhaustive search to obtain
optimal solutions in the general case. This has led some
researchers to consider the problem of PS as a combinatorial
optimization problem and use general techniques which are
known for their good results in similar situations.
Heuristics, especially the nearest neighbor algorithm, and
metaheuristics, especially genetic algorithms (GA) and tabu
search (TS) have been proposed to solve this problem.
III. IMPROVING PROTOTYPE SELECTION
Condensing algorithms try to find a significant reduction
of all prototypes so that the 1-NN classification gives results
as close as possible to those obtained using all the original
prototypes [7]. The problem that arises when this approach is
used is that it cannot provide a proof about the resulting sets
minimality. To correct this slight defect, the modern
heuristics or metaheuristics came to complete them. The
metaheuristics are strategies that guide the search towards an
optimal solution. These techniques are designed to explore the
search space efficiently in order to determine solutions
(almost) optimal. They may contain mechanisms to avoid
blocking in areas of space research.
Two types of metaheuristics have been successful in their
hybridization with the traditional KNN: genetic algorithms
and tabu search.
Genetic algorithms (GA) are an optimization technique
guided by the principles of natural evolution and genetics,
with a high presence of implicit parallelism. These algorithms
perform a search in complex, large, and multimodal
landscapes, and provide solutions for the quasi-optimal
objective function.
The tabu search (TS) is a method of dynamic
neighborhood, which selects, at each iteration, the best
solution of the first local optimum by finding the best
neighbor.
Vol. 2, No. 11, 2011
60 | P a g e
IV. HEURISTIC METHODS BASED ON THE CONDENSING K-
NN RULE
In 1968 Hart [17] was the first to propose a method
reducing the size of stored data for the nearest neighbor
decision. This method is called The Condensed Nearest
Neighbor Rule (CNN). The new idea about this rule
compared to the traditional KNN is the process of selecting a
subset TCNN from the initial training set TNN. TCNN must
be as effective as TNN in the classification of unknown
patterns. Actually the CNN rule minimizes the number of
models stored, keeping only a subset of training data for
classification, and employing a technique of low absorption.
The basic idea is to look for very similar training models, and
those that do not add additional information and eliminate
them.
This rule presents a case where TCNN consistency is not
achieved, when TNN is already the minimum set. In this case
TCNN will be equal to TNN, if this happens, the algorithm
will end if there are two equal models of different classes, but
TCNN must classify patterns correctly.
The technique that corrects this case of inconsistency is the
Reduced Nearest Neighbor rule (RNN) introduced by Gates
[11]. This rule is an extension of the CNN rule, and like CNN,
RNN reduces TNN.
The RNN algorithm starts at TRNN=TCNN and removes
every instance from TRNN if this deletion does not cause a
misclassification of another instance in TNN by the remaining
instances in TRNN. From the perspective of computing, it is
more expensive than the rule proposed by Hart, but it will
always produce a subset of CNN, and thus will be less
expensive in terms of computing and storage at the
classification stage.
Angiulli introduced the Fast Condensed Nearest
Neighbor rule (FCNN) [2] a scalable algorithm on large
multidimensional data sets used to create subsets serving as
consistent training sets based on the nearest neighbor decision
rule. This algorithm allows selecting points very close to the
border decision. It is independent of the order, and has low
quadratic complexity.
The FCNN rule initializes the consistent subset S with a
starting element from each class label of the training set.
The starting elements that the rule uses are particularly the
classs centroids in the training set. The algorithm is
incremental. At each iteration the resulting set is increased till
the stop condition is reached.
Wilson and Martinez suggested [21] a series of six
algorithms for sets reduction based on the kNN algorithm
where each algorithm improves the previous one. The first
reduction technique presented was the DROP1 which
constitutes the basic framework on which were built the five
other techniques. The DROP1 represents an improvement of
the RNN rule, which verifies the accuracy of the resulting set
instead of the initial set T. This algorithm is based on the
following rule: an instance P is removed only if at least some
of its associates (neighbors from same class) in S can be
classified correctly without P.
The first proposal causes a problem when the noisy
instances, which are typically associated with a different class,
cover only a small portion of the input space. DROP2 tries to
solve this problem by considering the effect of removing an
instance on all instances of the initial training T rather than S.
For this purpose the rule of DROP1 has been improved to
eliminate P if at least an acceptable number of its associates in
T can be classified correctly without P.
DROP2 sorts S in an attempt to remove the central points
before the border points (points which are near from the
decision boundaries), however noisy instances can also be on
borders, which can cause a change in the removal order. And
even if a noisy instance is central attempting to remove it
could eliminate border points that should be maintained.
Hence DROP3 uses a noise filtering before sorting the
instances of S. This is done using a rule that eliminates any
instance misclassified by its k nearest neighbors. Except that
DROP3 can sometimes remove an overly large number of
instances.
DROP4 improves DROP3 rule and provides that an
instance is removed only if:
1. It is misclassified by its k nearest neighbors, and
2. Its removal does not affect the classification of other
instances
DROP5 upgrades DROP2 by proposing that the instances
are considered beginning from the ones closest to the nearest
enemy (an enemy is the nearest neighbor of an instance with a
different class) and proceeding to outside.
The latest algorithm proposed by Wilson and Martinez
[22] was the DEL which is similar to DROP3, except that it
uses the length coding heuristic for deciding whether an
instance can be removed or not. In DEL an instance is
removed only if:
2. The removal of the instance does not increase the
cost of length encoding
Wu, Ianakiev and Govindraju proposed an Improved K-
Nearest Neighbor Classification [24]. A solution to increase
the speed of traditional kNN classification while maintaining
its level of accuracy by suggesting two building techniques.
The suggested IKNN algorithm is based on iterative
elimination of models with high attraction capacity.
In the first technique called the model condensing, the
authors suggested that all classes have the same probability,
and that the training set is initially created by the extraction
of vector elements from a large set of images of interest,
where each class has an equal representation. This set is
refined iteratively.
In the second technique, which is the pre-processing, an
unknown pattern is compared to a prototype in two stages. In
the first stage a rapid evaluation of the potential match is
made. The prototypes that fail in the first match are not
included in the second. Then, for a complete match, in the
Vol. 2, No. 11, 2011
61 | P a g e
second stage the norms difference of the prototype and the test
pattern must be less than a predetermined threshold.
The observation that large sets of data had computational
requirements which could be prohibitive to classify models
using the kNN, has led Fayed and Atia [9] to propose the
TRKNN, a way to lessen this problem through a condensation
approach. The aim of their approach is to eliminate the reasons
that makes load the calculation and does not contribute to
improve the classification.
This approach consists in rejecting the prototypes that are
far from the limits and have just a little influence on the KNN
classification. To achieve this, the authors first introduced the
concept of chain of nearest neighbors which is a sequence of
the nearest neighbors from alternating classes.
This technique proposes that if the distances of a given
model decrease in value, this model is considered to be
probably an internal point and can be discarded, whereas if the
distances do not decrease too much, then this point probably
varies around the limit of classification and it is maintained.
In a more recent paper [14] the authors introduced a new
approach "The Class Boundary Preserving Algorithm" (CBP),
a multi-step method for pruning the training set. The proposed
method aims at preserving instances that are close to the
borders of classes. Because, according to the authors, these
instances can provide most of the necessary information to
properly describe the underlying distribution. On the other
hand instances distant from limits are considered redundant by
the authors.
The innovation in this approach is the procedure used to
divide the training set into two subsets, X
B
containing the
instances near the surface of decision, and X
NB
containing the
internal samples. And because there is a noticeable difference
in the importance of information held by these two sets, two
different reduction processes were applied on both.
The algorithm considers an initial set of n instances related
to a set of labels, and involves four main steps.
The first step deals with the smoothing of class boundaries,
the second allows to distinguish between border and non-
border instances. The third step has to do with the pruning of
border instances and the last step is the clustering of non-
border instances.
To implement this idea, the authors have used the Mean
Shift Clustering algorithm (MSC) which converges to the
points of maximum density to determine the cluster centers of
the distribution. Then to make the maximum number of
neighbors in obtained clusters, the authors applied a merger
process which verifies the calculated clusters centers labels,
and if the clusters of nearest neighbors share the same label,
then the centers are merged.
A. Comparison
After examining some articles that have tackled the
reduction of sets based on the condensing KNN, we consider
useful to raise a comparative board grouping the idea, the
advantages, and disadvantages of each algorithm in the cited
articles.
TABLE I. COMPARISON OF NEAREST NEIGHBOR CONDENSING TECHNIQUES
CNN Idea Remove Training models that are very similar and those that do not add additional information to the
classification
Advantages - Improves search time and memory requirements
- Reduces the size of training data.
disadvantages - CNN is order dependent, then it is unlikely that it removes border points
- If the initial set is minimal, that causes an inconsistency in the resulting set when the program is
stopped.
- There is no guarantee that the resulting set is minimal.
RNN Idea Initially the resulting set is equal to the initial set, and then each instance that does not cause a wrong
classification of another instance in the initial set is removed from the resulting set.
Advantages - Reduces the size of training data and eliminates models
- Improves the search time and memory requirements.
Disadvantages - High computational cost
- Time consuming
- Its consistency depends on the consistency of the resulting set of CNN
FCNN Idea Select points very close to the decision boundary
Advantages - Average efficiency of 96.01% for an average number of iterations of about 69 iterations
- Has a smaller complexity than CNN
- Good rate of condensation
- Independent of the order
Disadvantages - Requires a large number of iterations
DROP1

Idea Initially the resulting set is equal to the initial set, and then an instance is removed only if at least some of
his associates in the resulting set can be ordered without it.
Advantages - Reduces the size of the training data and eliminates instances
- Constitutes a basis on which are built the rest of the DROP algorithms and DEL
- Noise do not degrade accuracy
Disadvantages - Checks the consistency of the resulting set instead of the initial set
- Low accuracy
- Cannot use information from previously eliminated instances
DROP2 Idea Improves DROP1 and eliminates an instance when at least a good number of its associates in the initial
set can be classified without it.
Advantages - Checks the consistency of the initial set rather than the final set
Vol. 2, No. 11, 2011
62 | P a g e
- Reduces the size of training data and eliminates instances
- Reach a higher accuracy than KNN for noisy instances
- Good storage reduction happens to 1 / 6 of the original set
- Storage requirements lower than DROP 1, 4, 5
Disadvantages - Attempts to remove the central points before the borderer, which overlooks the border noisy
instances
- The attempt to eliminate a noisy central instance can remove border points that should be
maintained.
- For very large data sets DROP2 does not eliminate enough points
DROP3 Idea Improves DROP2 by eliminating instances misclassified by their neighbors
Advantages - Is based on the classification of the instance itself to remove it
- Reach a higher accuracy than the traditional KNN in case of noisy instances
- Very good reduction of storage that comes to 12% of the original set
- Storage requirements lower than DROP1, 4, 5
Disadvantages - Can sometimes remove a too large number of instances
DEL Idea Improves DROP3 and eliminates an instance if:
2. The removal of the instance does not increase the cost of length encoding

Advantages - Achieved higher accuracy than the traditional KNN in case of noisy instances
Disadvantages - Accuracy less than that of DROP2-5
- Storage requirements higher than those of DROP2-5
IKNN Idea Remove iteratively models exhibitors of high capacity of attraction
Advantages - Reduces the data set by keeping the prototypes that are useful
- The preprocessing allows a significant saving in computation time
- The classifier shows a slight improvement in accuracy compared to traditional kNN
- Reduces the size of the models maintaining the same level of accuracy
Disadvantages - Classes must have same probability in the training set
- The work was built on intuition and not on a mathematical framework citing that the norm is an
intrinsic characteristic
- No theoretical proof that preprocessing guaranteed to filter the relevant prototypes or to maintain the
accuracy
- A test pattern is a distorted version of a prototype when the difference in standard is below a
threshold associates to this prototype
TRKNN Idea Eliminate the patterns that are a burden on computing and does not contribute to improve the
classification
Advantages - Reduces the size of the models without sacrificing accuracy
- TRKNN is up to 3 times faster than IKNN and up to 4 times than DROP2
Disadvantages - Average condensation rate (35%)
- Average level of accuracy
CBP Idea Preserve instances that are close to class boundaries
Advantages - Combine the selection and the abstraction
- Applies to each case an appropriate procedure of condensation
- The filtering Phase is based on the classification of the instance itself
- Use the geometric characteristics of the distribution
- Provides an overview of the distribution of samples
- Reduces the size of the training set and the number of representatives
- Good rate of condensation
Disadvantages - The filtering algorithm can sometimes remove a very large number of instances
- Relatively high computation time
These approaches which are based on traditional KNN,
select all the prototypes from training samples of initial set, by
adding or cutting, with aiming at preserving the performance
of classification with the use of heuristic methods. However,
none of these methods can prove the minimality of the result
set.
V. METAHEURISTICS: IMPROVING KNN
The basic concepts of metaheuristics can be described in
the abstract, without requiring a specific problem [15]. These
algorithms can therefore use heuristics, which in turn reflect
the specificity of the problem treated, except that these
heuristics are controlled by a higher level strategy.
In this article we present some improvements of classical
kNN algorithms through the use of some metaheuristics such
as genetic algorithms and tabu search.
R. Gil-Pita and X. Yao proposed [12] three improvements
of the k-nearest neighbor using genetic algorithms: the use of
an objective function based on mean square error, the
implementation of a clustered crossover, and a fast smart
mutation scheme.
In the first proposal they make use of a novel objective
function based on the MSE function. They consider the kNN
as a system with C outputs, so that each output is calculated
using a defined equation. Therefore, the C outputs of the kNN
Vol. 2, No. 11, 2011
63 | P a g e
system are approximations of the posterior probabilities of the
data.
In the second proposal the authors suggest a new crossover
scheme for the GA, denominated clustered crossover (CC), in
order to improve the determination of the best subset of the
training set that minimizes the selected objective function by
considering possible to determine the relationship between the
different bits of the bit stream.
In the third proposal R. Gil-Pita and X. Yao describe the
application of a mutation scheme that allows selecting the best
gene or group of genes to be changed, taking into account the
variations of the objective function with respect to each gene
for a given set. They design a fast method for evaluating the
error variation when each gene is changed, and they propose a
mutation strategy based on these variations of the objective
function. The authors denominate this mutation scheme as fast
smart mutation (FSM) [13], as it allows increasing the
effectiveness of the mutation stage in the genetic algorithm.
Compared to the use of a classic kNN classifier, obtained
results demonstrate the good accuracy of the proposed GA-
based technique.
The authors of A hybrid classification method of k-nearest
neighbor, Bayesian methods and genetic algorithm [16]
introduced a hybrid technique including KNN, genetic
algorithms and Bayesian method. This approach consists of
eight steps that begin with the application of the Bayesian
algorithm and then the generation of new data using the
genetic algorithm and the application of the K-nearest
neighbors method. And finally, several iterations of these
algorithms occur orchestrated by the genetic mechanism.
According to the authors, the method is useful on data sets that
have a small amount of data. It generates an unlimited number
of data that have similar characteristics to the original data,
and improves these data according to the proposed algorithm.
This method shows better classification performance with
respect to classic methods such as expectation maximization
algorithm [5] used to develop it. It is intended, according to its
authors, to hardware solutions with low cost based on
clustering, to noisy data classification, and to classification in
the data sets with little data.
To overcome the limitations of kNN, an improved version
of KNN, Genetic KNN (GKNN), was proposed by Suguna
and Thanushkodi [19]. A genetic algorithm is combined with
the k-nearest neighbor algorithm (KNN). In the proposed
method, using the genetic algorithm, k-number of samples are
selected for each iteration and the accuracy of the
classification is calculated as fitness. The greater accuracy is
recorded each time. Thus, it is not necessary to calculate the
similarities between all samples, or to consider the weight of
the category.
The performance of GKNN classifier was compared with
traditional KNN. The experiments and results show that the
GKNN not only reduces the complexity of the KNN, but it
improves the classification accuracy.
Ceveron and Ferri [3] proposed a method for the
prototypes selection for the nearest neighbor rule which aims
to obtain an optimal or close to optimal solution by presenting
a new approach based on tabu search. The particularity of the
proposed tabu search is that it uses the equations from
objective genetic algorithms. In this approach all possible
subsets of prototypes constitute the space of solutions.
Possible moves from a particular subset consist of adding or
removing each of the n initial prototypes. The attribute used to
declare the movement taboo is the prototype that is added or
deleted.
It is worth mentioning that the results obtained with TS
consistently improved the classical condensing techniques
previously published in the literature using the Iris database.
Wus article [22] provides a method of selecting items
based on the tabu search algorithm and KNN. First, the KNN
algorithm is used to generate the initial solution needed for the
tabu search. Then, the tabu algorithm is applied to obtain an
optimal subset of items. The KNN algorithm uses the
relevance of the elements to eliminate those redundant in large
networks data, and the subset obtained is the initial solution of
the tabu search algorithm.
This algorithm was tested via creating intrusion detection
model. The tests show that by using the feature selection
method proposed in this paper, the detection performance of
the intrusion detection system is effectively improved without
compromising detection accuracy, and the detection time and
accuracy performance of the system is much better than that of
the current feature selection methods. As a result the method
proposed in this paper has proved to be effective and feasible.
VI. CONCLUSION
In this paper we tried to present and compare sets
reduction techniques based on the principle of nearest
neighbor. These techniques are of type "condensing". Both are
improvements compared to basic KNN. These improvements
have been proposed by the authors to reduce the training set to
gain on speed and space efficiencies. To ensure the minimality
of this training set we presented some recent proposals using
metaheuristics to check the optimality of the resulting set of
some KNN reduction techniques.
Note that each technique is very effective in a specific area
and in special circumstances.
REFERENCES
[1] Chang Bao Rong, Naghibzadeh Mahmoud, Czarnowski Ireneusz, Danesh Malihe
and Danesh Mohaddesh, Data Clustering Based on an Efficient Hybrid of K-
Harmonic Means, PSO and GA, Transactions of Computational Collective
Intelligence, vol. IV, pp. 125-140, Springer, 2011.
[2] Fabrizio Angiulli, Fast condensed nearest neighbor rule, Technical report,
Proceedings of the 22
nd
International Conference on Machine Learning, Bonn,
Germany, 2005.
[3] Vicente Cerveron and Francesc J. Ferri, Another move toward the minimum
consistent subset: A tabu search approach to the condensed nearest neighbor rule,
IEEE Transactions on Systems, Man, and Cybernetics, vol. 31, pp. 408413, 2010.
[4] T. M. Cover and P. E. Hart, nearest neighbor pattern classification, IEEE
Transaction on Information Theory, vol. 13, pp. 2127, Jan 1967.
[5] F. Dallaert, The expectation maximization algorithm, Technical report,
College of Computing, Georgia Institute of Technology, February 2002.
[6] E. M. Elnahrawy, Log based chat room monitoring using text categorization: A
comparative study, University of Maryland.
[7] Pavel Paclik, Elzbieta Pekalska and Robert P.W. Duin, Prototype selection for
Vol. 2, No. 11, 2011
64 | P a g e
dissimilarity-based classifiers, Pattern Recognition, vol. 39, pp. 189208, 2006.
[8] F. Bajramovic, Frank Mattern, Nicholas Butko, and Joachim Denzler A
comparison of nearest neighbor search algorithms for generic object recognition,
LNCS 4179, ACIVS 2006, pp. 11861197.
[9] Hatem A. Fayed and Amir F. Atiya, A novel template reduction approach for the
K-nearest neighbor method, IEEE Transactions on Neural Networks, vol. 20 (5),
pp. 890896, May 2009.
[10] O. Kirmemis and G. Toker, Text categorization using k nearest neighbor
classification, Survey paper, Middle East Technical University.
[11] G.Gates, The reduced nearest neighbor rule. IEEE Transactions on Information
Theory, vol. 18, pp. 431433, 1972.
[12] R. Gil-Pita and X. Yao, Using a genetic algorithm for editing k-nearest neighbor
classifiers, IDEAL 2007, LNCS, vol. 4881, pp. 1141-1150, 2007.
[13] X Gil-Pita, R. Yao, Evolving edited k-nearest neighbor classifiers, International
Journal of Neural Systems, vol. 18 (6), pp. 459467, Dec 2008.
[14] Q.H. Wu, K. Nikolaidis and J.Y. Goulermas, A class boundary preserving
algorithm for data condensation, Pattern Recognition, vol. 44, pp. 704-715, 2011.
[15] Sean Luke, Essentials of Metaheuristics. Lulu, 2009.
[16] Mutlu Avci Mehmet Aci and Cigdem Inan, A hybrid classification method of k
nearest neighbor, bayesian methods and genetic algorithm, Expert Systems with
Applications, vol. 37, pp. 5061-5067, 2010.
[17] P. Hart, The condensed nearest neighbor rule, IEEE Transactions on Information
Theory, vol. 14, pp. 515516, 1968.
[18] Y. Wu Shizen, An algorithm for remote sensing image classification based on
artificial immune b-cell network, The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences, vol.
XXXVII, Part B6b, pp. 107-111, Beijing 2008.
[19] Suguna and Dr. K. Thanushkodi, An improved k-nearest neighbor classification
using genetic algorithm, International Journal of Computer Science Issues, vol. 7,
pp. 1821, 2010.
[20] V.Vaidehi, S. Vasuhi, R.Kayalvizhi, K.Mariammal, Raghuraman.M.B, Person
authentication using face recognition, Proceedings of the World Congress on
Engineering and Computer Science, 2008.
[21] Randall Wilson and Tony R. Martinez, Reduction techniques for instance-based
learning algorithms, Machine Learning, vol. 38 (3), pp. 257286, 2000.
[22] Tao Ran, Wu Jian-guang and Li Zhi-Yong, An improving tabu search algorithm
for intrusion detection, IEEE Computer Society, pp. 435439, 2011.
[23] Y. Yang and T. Ault, Improving text categorization methods for event tracking,
Carnegie Mellon University.
[24] Krasimir G. Iankiev Yingquan Wu and Venu Govindaraju, Improved k-nearest
neighbor classification, Pattern recognition, vol. 35, pp. 23112318, 2002.
[25] A. V. Zukhba, Np-completeness of the problem of prototype selection
in the nearest neighbor method, Pattern Recognition and Image
Analysis, vol. 20, pp. 484-494, 2010.

Vol. 2, No. 11, 2011
65 | P a g e
Concurrent Edge Prevision and Rear Edge Pruning
Approach for Frequent Closed Itemset Mining

Anurag Choubey
1
Dean Academic, Technocrats
Institute of Technology,
RGPV,
Bhopal (M.P.), India

Dr. Ravindra Patel
2
Reader & Head, Dept. of Computer
Application,
UIT-RGPV,

Dr. J.L. Rana
3

Former Professor & Head, Dept. of
CSE & IT MANIT,
Abstract- Past observations have shown that a frequent item set
mining algorithm are purported to mine the closed ones because
the finish provides a compact and a whole progress set and
higher potency. Anyhow, the newest closed item set mining
algorithms works with candidate maintenance combined with
check paradigm that is pricey in runtime yet as space usage
when support threshold is a smaller amount or the item sets gets
long. Here, we show, CEG&REP that could be a capable
algorithm used for mining closed sequences while not candidate.
It implements a completely unique sequence finality verification
model by constructing a Graph structure that build by an
approach labeled Concurrent Edge Prevision and Rear Edge
Pruning briefly will refer as CEG&REP. a whole observation
having sparse and dense real-life knowledge sets proved that
CEG&REP performs bigger compared to older algorithms
because it takes low memory and is quicker than any algorithms
those cited in literature frequently.
Keywords- Data mining; Closed Itemsets; Pattern Mining;
sequence length; graph structure.
I. INTRODUCTION
Sequential item set mining, is an important task, having
many applications with market, customer and web log analysis,
item set discovery in protein sequences. Capable mining
techniques are being observed extensively, including the
general sequential item set mining [1, 2, 3, 4, 5, 6], constraint-
based sequential item set mining [7, 8, 9], frequent episode
mining [10], cyclic association rule mining [11], temporal
relation mining [12], partial periodic pattern mining [13], and
long sequential item set mining [14]. Recently its quite
convincing that for mining frequent item sets, one should mine
all the closed ones as the end leads to compact and complete
result set having high efficiency [15, 16, 17, 18], unlike mining
frequent item sets, there are less methods for mining closed
sequential item sets. This is because of intensity of the problem
and CloSpan is the only variety of algorithm [17], similar to the
frequent closed item set mining algorithms, it follows a
candidate maintenance-and-test paradigm, as it maintains a set
of readily mined closed sequence candidates used to prune
search space and verify whether a recently found frequent
sequence is to be closed or not. Unluckily, a closed item set
mining algorithm under this paradigm has bad scalability in the
number of frequent closed item sets as many frequent closed
item sets (or just candidates) consume memory and leading to
high search space for the closure checking of recent item sets,
which happens when the support threshold is less or the item
sets gets long.
Finding a way to mine frequent closed sequences without
the help of candidate maintenance seems to be difficult. Here,
we show a solution leading to an algorithm, CEG&REP, which
can mine efficiently all the sets of frequent closed sequences
through a sequence graph protruding approach. In CEG&REP,
we need not eye down on any historical frequent closed
sequence for a new patterns closure checking, leading to the
proposal of Sequence graph edge pruning technique and other
kinds of optimization techniques.
The observations display the performance of the
CEG&REP to find closed frequent itemsets using Sequence
Graph: The comparative study claims some interesting
performance improvements over BIDE and other frequently
cited algorithms.
In section II most frequently cited work and their limits
explained. In section III the Dataset adoption and formulation
explained. In section IV, introduction to CEG&REP and its
utilization for Sequence Graph protruding explained. In section
V, the algorithms used in CEG&REP described. In section V1,
results gained from a comparative study briefed and fallowed
by conclusion of the study.
II. RELATED WORK
The sequential item set mining problem was initiated by
Agrawal and Srikant , and the same developed a filtered
algorithm, GSP [2], basing on the Apriori property [19]. Since
then, lots of sequential item set mining algorithms are being
developed for efficiency. Some are, SPADE [4], PrefixSpan
[5], and SPAM [6]. SPADE is on principle of vertical id-list
format and it uses a lattice-theoretic method to decompose the
search space into many tiny spaces, on the other hand
PrefixSpan implements a horizontal format dataset
representation and mines the sequential item sets with the
pattern-growth paradigm: grow a prefix item set to attain
longer sequential item sets on building and scanning its
database. The SPADE and the PrefixSPan highly perform GSP.
SPAM is a recent algorithm used for mining lengthy sequential
item sets and implements a vertical bitmap representation. Its
observations reveal, SPAM is better efficient in mining long
Vol. 2, No. 11, 2011
66 | P a g e
item sets compared to SPADE and PrefixSpan but, it still takes
more space than SPADE and PrefixSpan. Since the frequent
closed item set mining [15], many capable frequent closed item
set mining algorithms are introduced, like A-Close [15],
CLOSET [20], CHARM [16], and CLOSET+ [18]. Many such
algorithms are to maintain the ready mined frequent closed
item sets to attain item set closure checking. To decrease the
memory usage and search space for item set closure checking,
two algorithms, TFP [21] and CLOSET+2, implement a
compact 2-level hash indexed result-tree structure to keep the
readily mined frequent closed item set candidates. Some
pruning methods and item set closure verifying methods,
initiated the can be extended for optimizing the mining of
closed sequential item sets also. CloSpan is a new algorithm
used for mining frequent closed sequences [17]. It goes by the
candidate maintenance-and-test method: initially create a set of
closed sequence candidates stored in a hash indexed result-tree
structure and do post-pruning on it. It requires some pruning
techniques such as Common Prefix and Backward Sub-Item set
pruning to prune the search space as CloSpan requires
maintaining the set of closed sequence candidates, it consumes
much memory leading to heavy search space for item set
closure checking when there are more frequent closed
sequences. Because of which, it does not scale well the number
of frequent closed sequences. BIDE [26] is another closed
pattern mining algorithm and ranked high in performance when
compared to other algorithms discussed. Bide projects the
sequences after projection it prunes the patterns that are subsets
of current patterns if and only if subset and superset contains
same support required. But this model is opting to projection
and pruning in sequential manner. This sequential approach
sometimes turns to expensive when sequence length is
considerably high. In our earlier literature[27] we discussed
some other interesting works published in recent literature.
Here, we bring Sequence Graph protruding that based on
edge projection and pruning, an asymmetric parallel algorithm
for finding the set of frequent closed sequences. The giving of
this paper is:
(A) an improved sequence graph based idea is generated for
mining closed sequences without candidate maintenance,
termed as Concurrent Edge Prevision and Rear Edge Pruning
(CEG&REP) based Sequence Graph Protruding for closed
itemset mining. The Edge Projection is a forward approach
grows till edge with required support is possible during that
time the edges will be pruned. During this pruning process
vertices of the edge that differs in support with next edge
projected will be considered as closed itemset, also the
sequence of vertices that connected by edges with similar
support and no projection possible also be considered as closed
itemset
(B) in the Edge Projection and pruning based Sequence
Graph Protruding for closed itemset mining, we create a
algorithms for Edge Prevision and Rear Edge Pruning
(C) The performance clearly signifies that proposed model
has a very high capacity: it can be faster than an order of
magnitude of CloSpan but uses order(s) of magnitude less
memory in several cases. It has a good scalability to the
database size. When compared to BIDE the model is proven as
equivalent and efficient in an incremental way that proportional
to increment in pattern length and data density.
III. DATASET ADOPTION AND FORMULATION
Item Sets I: A set of diverse elements by which the
sequences generate.
1
n
k
k
I i
=
=
U
Note: I is set of diverse elements

Sequence set S: A set of sequences, where each sequence
contains elements each element e belongs to I and true for a
function p(e). Sequence set can formulate as

Represents a sequences of items those belongs to set of
distinct items I.
m: total ordered items.
P(e
i
): a transaction, where e
i
usage is true for that transaction.

S: represents set of sequences
t: represents total number of sequences and its value is
volatile
s
j
: is a sequence that belongs to S
Subsequence: a sequence of sequence set S is
considered as subsequence of another sequence
of Sequence
Set S if all items in sequence S
p
is belongs to s
q
as an ordered
list. This can be formulated as
If
Then
where

Total Support ts : occurrence count of a sequence as an
ordered list in all sequences in sequence set S can adopt as
total support ts of that sequence. Total support ts of a
sequence can determine by fallowing formulation.

Is set of sequences
: Represents the total support ts of sequence s
t
is
the number of super sequences of s
t

1
| ( ( ), )
m
i i i
i
s e p e e I
=
= < e >
U
1
t
j
j
S s
=
=
U
p
s
q
s
1
( ) ( )
n
pi q p q
i
s s s s
=
e _
U
1 1
:
n m
pi qj
i j
s s
= =
<
U U

p q
s S and s S e e
( ) | : ( 1.. | |) |
ts t t p S
f s s s for each p DB = < =
S
DB
( )
ts t
f s
Vol. 2, No. 11, 2011
67 | P a g e
Qualified support q
s
: The resultant coefficient of total
support divides by size of sequence database adopt as qualified
support qs. Qualified support can be found by using fallowing
formulation.

Sub-sequence and Super-sequence: A sequence is sub
sequence for its next projected sequence if both sequences
having same total support.
Super-sequence: A sequence is a super sequence for a
sequence from which that projected, if both having same total
support.
Sub-sequence and super-sequence can be formulated as
If rs where rs is required support threshold given
by user
And where

IV. CONCURRENT EDGE PREVISION AND REAR EDGE
PRUNING
A. Preprocess:
As a first stage of the proposal we perform dataset
preprocessing and itemsets Database initialization. We find
itemsets with single element, in parallel prunes itemsets with
single element those contains total support less than required
support.
B. Edge Prevision:
In this phase, we select all itemsets from given itemset
database as input in parallel. Then we start projecting edges
from each selected itemset to all possible elements. The first
iteration includes the pruning process in parallel, from second
iteration onwards this pruning is not required, which we
claimed as an efficient process compared to other similar
techniques like BIDE. In first iteration, we project an itemset
that spawned from selected itemset from and an
element considered from I. If the is greater or
equal to , then an edge will be defined between and . If
then we prune from . This pruning
process required and limited to first iteration only.
From second iteration onwards project the itemset that
spawned from to each element of I. An edge can be
defined between and if is greater or equal to
. In this description is a projected itemset in previous
iteration and eligible as a sequence. Then apply the fallowing
validation to find closed sequence.
If any of that edge will be pruned and
all disjoint graphs except will be considered as closed
sequence and moves it into

and remove all disjoint
graphs from memory.
If and there after no projection
spawned then will be considered as closed sequence and
moves it into and remove from memory.
The above process continues till the elements available in
memory those are connected through direct or transitive edges
and projecting itemsets i.e., till graph become empty
1) Algorithm used in CEG&REP:
This section describes algorithms for initializing sequence
database with single elements sequences, spawning itemset
projections and pruning edges from Sequence Graph SG.
Algorithm 1: Concurrent Edge Prevision to build graph
structure and Rear Edge Pruning
Step 1:
Input: Set of Elements I.
Begin:
L1: For each element of I
Begin:
Find
If then
Move as sequence with single element to
End: L1.
End.
Step 2:
Input: and I;
L1: For each sequence in
Begin:
L2: For each element of I
Begin:
C1:
Begin:
Create projected itemset from
If then prune from
End: C1.
End: L2.
End: L1.
( )
( )
| |
ts t
qs t
S
f s
f s
DB
=
( )
ts t
f s >
:
t p
s s for any pvalue < ( ) ( )
ts t ts p
f s f s ~
p
s
i
s
S
DB
i
e ( )
ts p
f s
rs
i
s
i
e
( ) ( )
ts i ts p
f s f s ~
i
s
S
DB
p
S
' p
S
i
e
' p
S
i
e ( )
ts p
f s
rs
' p
S
'
( ) ( )
ts p ts p
f s f s ~
p
s
S
DB
'
( ) ( )
ts p ts p
f s f s ~
p
s
S
DB
'

p p
s and s
i
e
( )
ts i
f e
( )
ts i
f e rs >
i
e
S
DB
S
DB
i
s
S
DB
i
e
if edgeWeight( , )
i i
s e rs >
p
s ( , )
i i
s e
( ) ( )
ts i ts p
f s f s ~
i
s
S
DB
Vol. 2, No. 11, 2011
68 | P a g e
L3: For each projected Itemset in memory
Begin:

L4: For each of I
Begin:
Project from
C2: If
Begin
Spawn SG by adding edge between
End: C2
End: L4
C3: If not spawned and no new projections added for

Begin:
Remove all duplicate edges for each edge weight from
and keep edges unique by not deleting most recent edges for
each edge weight.
Select elements from each disjoint graph as closed
sequence and add it to and remove disjoint graphs from
SG.
End C3
End: L3
If go to L3

Fig 1: Concurrent Edge Prevision to build graph structure and Rear Edge Pruning

VI. COMPARATIVE STUDY:
This segment focuses mainly on providing evidence on
asserting the claimed assumptions that 1) The CEG&REP is
similar to BIDE which is actually a sealed series mining
algorithm that is competent enough to momentously surpass
results when evaluated against other algorithms such as
CloSpan and spade. 2) Utilization of memory and momentum
is rapid when compared to the ColSpan algorithm which is
again analogous to BIDE. 3) There is the involvement of an
enhanced occurrence and a probability reduction in the
memory exploitation rate with the aid of the trait equivalent
prognosis and also rim snipping of the CEG&REP. This is on
the basis of the surveillance done which concludes that
CEG&REPs implementation is far more noteworthy and
important in contrast with the likes of BIDE, to be precise.
p
s
' p p
s s =
i
e
p
s
'
( , )
p i
s e
( )
ts p
f s rs >
'

p i
s and e
' p
s
' p
s
' p
s
S
DB
SG | =
Vol. 2, No. 11, 2011
69 | P a g e
JAVA 1.6_ 20th build was employed for accomplishment
of the CEG&REP and BIDE algorithms. A workstation
equipped with core2duo processor, 2GB RAM and Windows
XP installation was made use of for investigation of the
algorithms. The parallel replica was deployed to attain the
thread concept in JAVA.
A. Dataset Characteristics:
Pi is supposedly found to be a very opaque dataset, which
assists in excavating enormous quantity of recurring clogged
series with a profitably high threshold somewhere close to
90%. It also has a distinct element of being enclosed with 190
protein series and 21 divergent objects. Reviewing of
serviceable legacys consistency has been made use of by this
dataset. Fig. 5 portrays an image depicting dataset series extent
status.
In assessment with all the other regularly quoted forms like
spade, prefixspan and CloSpan, BIDE has made its mark as a
most preferable, superior and sealed example of mining copy,
taking in view the detailed study of the factors mainly, memory
consumption and runtime, judging with CEG&REP.

Fig 3: A comparison report for Runtime

Fig: 4: A comparison report for memory usage
In contrast to CEG&REP and BIDE, a very intense dataset
Pi is used which has petite recurrent closed series whose end to
end distance is less than 10, even in the instance of high
support amounting to around 90%. The diagrammatic
representation displayed in Fig 3 explains that the above
mentioned two algorithms execute in a similar fashion in case
of support being 90% and above. But in situations when the
support case is 88% and less, then the act of CEG&REP
surpasses BIDEs routine.

Fig 5: Sequence length and number of sequences at different thresholds in Pi
dataset
The disparity in memory exploitation of CEG&REP and
BIDE can be clearly observed because of the consumption
level of CEG&REP being low than that of BIDE.
V. CONCLUSION
It has been scientifically and experimentally proved that
clogged prototype mining propels dense product set and
considerably enhanced competency as compared to recurrent
prototype of mining even though both these types project
similar animated power. The detailed study has verified that the
case usually holds true when the count of recurrent moulds is
considerably large and is the same with the recurrent bordered
models as well. However, there is the downbeat in which the
earlier formed clogged mining algorithms depend on
chronological set of recurrent mining outlines. It is used to
verify whether an innovative recurrent outline is blocked or
else if it can nullify few previously mined blocked patterns.
This leads to a situation where the memory utilization is
considerably high but also leads to inadequacy of increasing
seek out space for outline closure inspection. This paper
anticipates an unusual algorithm for withdrawing recurring
closed series with the help of Sequence Graph. It performs te
following functions: It shuns the blight of contenders
maintenance and test exemplar, supervises memory space
expertly and ensures recurrent closure of clogging in a well-
organized manner and at the same instant guzzling less amount
of memory plot in comparison with the earlier developed
mining algorithms. There is no necessity of preserving the
already defined set of blocked recurrences, hence it very well
balances the range of the count of frequent clogged models. A
Sequence graph is embraced by CEG&REP and has the
capability of harvesting the recurrent clogged pattern in an
online approach. The efficacy of dataset drafts can be
showcased by a wide-spread range of experimentation on a
number of authentic datasets amassing varied allocation
attributes. CEG&REP is rich in terms of velocity and memory
spacing in comparison with the BIDE and CloSpan algorithms.
ON the basis of the amount of progressions, linear scalability is
provided. It has been proven and verified by many scientific
research studies that limitations are crucial for a number of
chronological outlined mining algorithms. Future studies
include proposing of claiming a deduction advance on perking
up the rule coherency on predictable itemsets.

Vol. 2, No. 11, 2011
70 | P a g e
REFERENCES
[1] F. Masseglia, F. Cathala, and P. Poncelet, The psp approach for mining
sequential patterns. In PKDD98, Nantes, France, Sept. 1995.
[2] R. Srikant, and R. Agrawal, Mining sequential patterns: Generalizations
and performance improvements. In EDBT96, Avignon, France, Mar.
1996.
[3] J. Han, J. Pei, B. Mortazavi-Asl, Q. Chen, U. Dayal, and M.C. Hsu,
FreeSpan: Frequent pattern-projected sequential pattern mining . In
SIGKDD00, Boston, MA, Aug. 2000.
[4] M. Zaki, SPADE: An Efficient Algorithm for Mining Frequent
Sequences. Machine Learning, 42:31-60, Kluwer Academic Pulishers,
2001.
[5] J. Pei, J. Han, B. Mortazavi-Asl, Q. Chen, U. Dayal, and M.C. Hsu,
PrefixSpan: Mining sequential patterns efficiently by prefix-projected
pattern growth. In ICDE01, Heidelberg, Germany, April 2001.
[6] J. Ayres, J. Gehrke, T. Yiu, and J. Flannick, Sequential PAttern Mining
using a Bitmap Representation. In SIGKDD02, Edmonton, Canada, July
2002.
[7] M. Garofalakis, R. Rastogi, and K. Shim, SPIRIT: Sequential PAttern
Mining with regular expression constraints. In VLDB99, San Francisco,
CA, Sept. 1999.
[8] J. Pei, J. Han, and W. Wang, Constraint-based sequential pattern mining
in large databases. In CIKM02, McLean, VA, Nov. 2002.
[9] M. Seno, G. Karypis, SLPMiner: An algorithm for finding frequent
sequential patterns using lengthdecreasing support constraint. In
ICDM02,, Maebashi, Japan, Dec. 2002.
[10] H. Mannila, H. Toivonen, and A.I. Verkamo, Discovering frequent
episodes in sequences . In SIGKDD95, Montreal, Canada, Aug. 1995.
[11] B. Ozden, S. Ramaswamy, and A. Silberschatz, Cyclic association rules.
In ICDE98, Olando, FL, Feb. 1998.
[12] C. Bettini, X. Wang, and S. Jajodia, Mining temporal relationals with
multiple granularities in time sequences. Data Engineering Bulletin,
21(1):32-38, 1998.
[13] J. Han, G. Dong, and Y. Yin, Efficient mining of partial periodic patterns
in time series database. In ICDE99, Sydney, Australia, Mar. 1999.
[14] J. Yang, P.S. Yu, W. Wang and J. Han, Mining long sequential patterns in
a noisy environment. In SIGMOD 02, Madison, WI, June 2002.
[15] N. Pasquier, Y. Bastide, R. Taouil and L. Lakhal, Discoving frequent
closed itemsets for association rules. In ICDT99, Jerusalem, Israel, Jan.
1999.
[16] M. Zaki, and C. Hsiao, CHARM: An efficient algorithm for closed
itemset mining. In SDM02, Arlington, VA, April 2002.
[17] X. Yan, J. Han, and R. Afshar, CloSpan: Mining Closed Sequential
Patterns in Large Databases. In SDM03, San Francisco, CA, May 2003.
[18] J. Wang, J. Han, and J. Pei, CLOSET+: Searching for the Best Strategies
for Mining Frequent Closed Itemsets. In KDD03, Washington, DC, Aug.
2003.
[19] R. Agrawal and R. Srikant. Fast algorithms for mining association rules.
In VLDB94, Santiago, Chile, Sept. 1994.
[20] J. Pei, J. Han, and R. Mao, CLOSET: An efficient algorithm for mining
frequent closed itemsets . In DMKD01 workshop, Dallas, TX, May
2001.
[21] J. Han, J. Wang, Y. Lu, and P. Tzvetkov, Mining Top- K Frequent Closed
Patterns without Minimum Support. In ICDM02, Maebashi, Japan, Dec.
2002.
[22] P. Aloy, E. Querol, F.X. Aviles and M.J.E. Sternberg, Automated
Structure-based Prediction of Functional Sites in Proteins: Applications to
Assessing the Validity of Inheriting Protein Function From Homology in
Genome Annotation and to Protein Docking. Journal of Molecular
Biology, 311, 2002.
[23] R. Agrawal, and R. Srikant, Mining sequential patterns. In ICDE95,
Taipei, Taiwan, Mar. 1995.
[24] I. Jonassen, J.F. Collins, and D.G. Higgins, Finding flexible patterns in
unaligned protein sequences. Protein Science, 4(8), 1995.
[25] R. Kohavi, C. Brodley, B. Frasca, L.Mason, and Z. Zheng, KDD-cup
2000 organizers report: Peeling the Onion. SIGKDD Explorations, 2,
2000.
[26] Jianyong Wang, Jiawei Han: BIDE: Efficient Mining of Frequent Closed
Sequences. ICDE 2004: 79-90
[27] Anurag Choubey, Dr. Ravindra Patel and,Dr. J.L. Rana. Article: Frequent
Pattern Mining With Closeness Considerations: Current State Of The Art.
GJCST Issue 11, Volume 17, August 2011. Published by Global Journals,
25200 Carlos Bee Blvd. #495, Hayward, CA 94542, USA
AUTHORS PROFILE
Anurag Choubey
1
: He has completed B.Sc. (Electronics) in 1990 from Dr.
Hari Singh Gaur Vishwavidyalaya, Sagar, M.P., India and M.Sc. (Applied
Physics) in 1993 from Govt. Engineering College, Jabalpur, M.P.,
India. He has completed MCA in 2008 from Guru Ghasidas
Vishwavidyalaya (Presently central University), Bilaspur C.G.,
India. He has worked as a Lecturer at Govt. Engg. College,
Jabalpur and Hitkarni College of Engineering & Technology,
Jabalpur from 1998 to 2000. From October 2000 onwards he has
been working in Technocrats Institute of Technology, Bhopal. Currently
working as Dean Academic, Technocrats Institute of Technology, Bhopal M.P.,
India. He posses more than 13 years of experience in teaching and has worked
in different capacities like controller of exam, admission in-charge and other
administrative post. email: anuragphd11@gmail.com
Dr. Ravindra Patel
2
: Associate Professor and Head, Department of Computer
Applications at Rajiv Gandhi Technological University, Bhopal, India. He has
awarded Ph.D. degree in Computer Science. He possessed more
than 10 years of experience in teaching post-graduate classes. He
has published more than 15 papers in international and national
journals and conference proceedings. He is member of
International Association of Computer Science and Information
Technology (IACSIT). Email: ravindra@rgtu.net
Dr. J.L. Rana
3
: Former Professor and Head, Department of
Computer Science and Engineering & Information Technology
at MANIT, Bhopal, India with 32 years of vast experience of
teaching, out of which 19 years as Professor of Computer
Science & Engg. and Information Technology. Currently he is
working as Group Director of Radha Raman Group of
Institute, Bhopal. He has completed B.E. (Hons) in 1968 from GEC, Jabalpur
and M.S. Computer Control from University of Hawaii (UH), USA in 1972. He
has awarded Ph.D. degree in Computer from Indian Institute of Technology,
Bombay in 1987. He also posses more than 25 years of experience in post
graduate teaching He has published more than 30 papers in international and
national journals and 65 international and national conference proceedings. He
has guided 12 nos. of Ph.D. and 5 no. in progress He is a senior life member of
CSI and Chairman CSI Bhopal Chapter for two terms. Email:
jl_rana@yahoo.com

Vol. 2, No. 11, 2011
71 | P a g e
Error Filtering Schemes for Color Images in Visual
Cryptography

Shiny Malar F.R
Dept. of Computer Science & Engineering
Noorul Islam University, Kumaracoil
Kanyakumari district, India
Jeya Kumar M.K
Professor, Dept. of Computer Applications
Noorul Islam University, Kumaracoil
Kanyakumari District, India

Abstract - The color visual cryptography methods are free from
the limitations of randomness on color images. The two basic
ideas used are error diffusion and pixel synchronization. Error
diffusion is a simple method, in which the quantization error at
each pixel level is filtered and fed as the input to the next pixel. In
this way low frequency that is obtained between the input and
output image is minimized which in turn give quality images.
Degradation of colors are avoided with the help of pixel
synchronization. The proposal of this work presents an efficient
color image visual cryptic filtering scheme to improve the image
quality on restored original image from visual cryptic shares. The
proposed color image visual cryptic filtering scheme presents a
deblurring effect on the non-uniform distribution of visual
cryptic share pixels. After eliminating blurring effects on the
pixels, Fourier transformation is applied to normalize the
unevenly transformed share pixels on the original restored image.
This in turn improves the quality of restored visual
cryptographic image to its optimality. In addition the overlapping
portions of the two or multiple visual cryptic shares are filtered
out with homogeneity of pixel texture property on the restored
original image. Experimentation are conducted with standard
synthetic and real data set images, which shows better
performance of proposed color image visual cryptic filtering
scheme measured in terms of PSNR value (improved to 3 times)
and share pixel error rate (reduced to nearly 11%) with existing
grey visual cryptic filters. The results showed that the noise
effects such as blurring on the restoration of original image are
removed completely.
Keywords - Error Diffusion; Visual Cryptography; Fourier
Filtering; Context Overlapping; Color Extended Visual
Cryptography.
I. INTRODUCTION
Visual Cryptography, an encryption technique allows
cryptic to be possible only if the proper key is supplied by the
user and decryption can be performed without the intervention
of the computer. It works on the principle that when an image
is splited into k shares only the user who has all the k shares
can decrypt the message, any k-1 shares held by the user do
not contain any useful information[1].
Naor and Shamir [2], in 1994 proposed a new security
technique named visual cryptography scheme. In this
technique, a secret image of type binary is encoded in a
cryptographical manner into random binary patterns which
contains n shares in a k-out-of-n scheme. The n shares are
distributed among n participants in such a way the each
participants share is not known to another participant. The
secret image can be visually revealed by k or more participants
by joining all the shares available. Even if computational
power decoding is available, cannot be done on the secret
image by k-1 or fewer participants.
As the shares in the layers occur as random noise, the
attackers cannot identify any useful information about the
individual shares. Even with the availability of computer, it is
not possible to decrypt the message or information with the
limited availability of the share. The limitation of the above
method is its randomness without any visual information.
Extended Visual Cryptography have been suggested which
also suffers from the same drawbacks of randomness. This
paper is well thought-out as follows, Section II deals with the
review of literature. Section III described about the error
filtering schemes for color images. Section IV and V offered
to Experimental result and discussion .Finally the conclusion
of this paper in Section VI.
II. RELATED WORKS
Recently in the literature, many new methods have been
implemented for visual cryptography. In 1995 Naor and
Shamir [3], have predicted an optimal dissimilarity in k-out-
of-n scheme to alleviate the contrast loss problem in the
reconstructed image. A visual cryptography scheme is a broad
spectrum method which is based upon general access
structure. In k-out-of-n secret sharing scheme, any k shares
will decode the secret image, which reduce the security level.
To overcome this problem the basic secret sharing scheme is
extended to general access structure. The concept of general
access structure method was introduced in the year 1996 and
1997, by Ateniese, C.Blundo, A.Desantis and D.R.Stinson [4 ,
5,6,7].In 1999,[8,9] Image size invariant visual cryptography
was introduced by R. Ito, H. Kuwakado, and H. Tanaka and
also in the same year the C.-N. Yang and C.-S. Laih have
proposed some new types of visual secret sharing schemes.
In previous works of visual cryptography, binary images
were concentrated which is not enough in real time
applications. This general access structure method is applied to
the gray level images are introduced by L. A.
MacPherson,Chang Choulin[10,11,12],in the year 2000. In
2001 the G. Ateniese, C. Blundo, A. Santis, and D. R. Stinson
have predicted the extended capabilities for visual
cryptography in the natural images [13-16]. Ateniese has
projected the hypergraph coloring method for Visual
Vol. 2, No. 11, 2011
72 | P a g e
cryptography , which is used to construct meaningful binary
shares. Since hypergraph colorings are constructed by random
distributed pixels, this method produce insufficient results. A
new method of Extended visual cryptography for natural
images is used to produce meaningful binary shares which is
predicted by Nakajima[17,18] in the year 2002.Wen-Hsiang
Tsai[19-23] have estimated the dithering technique which is
applied to gray level images in visual cryptography. This
technique is used for transformation of gray level images into
binary images in the year 2003 . Again, Hou[24,25] has
proposed the binary visual cryptography scheme which is
applied to gray level images, that a gray level image is
converted into halftone images in the year 2004.
In 2006 the Zhi Zhou, Gonzalo, R.Arce and Giovanni
Dicrescenzo [29-33] have proposed halftone visual
cryptography which produce good quality and meaningful
halftone shares, the generated halftone shares contain the
visual information. In halftone visual cryptography a secret
binary pixel P is encoded into an array of Q1 x Q2 (m in
basic model) sub pixels, referred to as halftone cell in each of
the n shares. By using halftone cells with an appropriate size,
visually pleasing halftone shares can be obtained and also
maintained contrast and security. Abhishek parakh and
Subhash Kak have proposed recursive threshold visual
cryptography which is used in network applications and also
reduce the network load. In 2007 the C.M. Hu and W.G.
Tzeng [34, 35] have proposed a cheating method in Visual
Cryptography schemes. In their cheating method, the cheater
needs to know the exact distribution of black and white sub
pixels of the shares of honest participants. In the same year, a
Cheating Prevention Scheme for Binary Visual Cryptography
with Homogeneous Secret Images was introduced by D.S.
Tsai, T.H. Chen, G. Horng, Which is used to prevent the
cheater from obtaining the distribution [26, 27, 28].
However, the knowledge of distribution is not a necessary
condition for a successful cheats. They also proposed another
cheat-preventing method in which the stacking of the genuine
share and verification share reveals the verification image in
some small region that it is possible to attack the method.
Niranjan Damera-Venkata, and Brian L. Evans have predicted
the design and analysis of vector color error diffusion
halftoning systems. And also quantization of accumulate
diffused errors in error diffusion method was introduced by Ti-
Chiun Chang and Jan P. Allebach in the year 2005 [26, 27,
28].
In 2009 the Zhongmin Wang, Gonzalo R. Arce,, and
Giovanni Di Crescenzo [36,37] have proposed the Visual
Cryptography for color image using visual information pixel
(VIP) synchronization with error diffusion technique. They
are introduced a color Visual Cryptography encryption method
which leads to significant shares and is free of the previously
mentioned limitations. This method is used to filtering the
error in an image and produces the meaningful shares. The
error filtering schemes for color images is very simple and
efficient method.
III. ERROR FILTERING SCHEMES FOR COLOR IMAES
A. Fourier filtering for color visual cryptographic images
The Fourier Transform of an image can be carried out
using the Discrete Fourier Transform (DFT) method. Fig.1
shows the DFT also allows spectral data (i.e. a transformed
image) to be inverse transformed, producing an image once
again. If we compute the DFT of an image, then immediately
inverse transform the result, we expect to regain the same
image. If we multiply each element of the DFT of an image by
a suitably chosen weighting function we can accentuate certain
frequency components and attenuate others. The
corresponding changes in the spatial form can be seen after the
inverse DFT has been computed.
Figure 1. Fourier filtering for color visual cryptographic images.
The selective enhancement/suppression of frequency
components is known as Fourier Filtering. The fourier filtering
is used for convolution with large masks (Convolution
Theorem), compensate for known image defects (restoration),
reduction of image noise, suppression of hum or other
periodic interference and reconstruction of original restored
visual cryptographic image.
1) Fourier Filtering
The DFT is the sampled Fourier Transform and does not
have all frequencies to form an image, but only a set of forms
which is large enough to fully define the spatial domain image.
The total number of frequencies correspond to the total
number of pixels in the spatial domain image, i.e. the image
in the spatial and Fourier domain is of the equal size.
For a square image of size NN, the two-dimensional DFT
is shown in the equation 1.
Ima
ge
Inverse
transform

DFT
metho
d
DFT
image

Fouri
er
filteri
ng
Invers
e DFT
Original
VC Image
Frequency
component
s
Multiply with
weighting
function

Spectral
data

Vol. 2, No. 11, 2011
73 | P a g e
Input
image
Input
image
Texture
overlapping
filters
Overlapped
texture
Optimal patch
region
Max-
flow/min-
cut
algorithm
Perceptual
effect by MRF
method
N-1 N-1
F ( x,y ) = f (p,q) e
-y2 ( xp/N + yq/N )

(1)
p=0 q=o

where f(p,q) is the image in the spatial domain . The
exponential term is the basic function corresponding to each
point F(x,y)in the Fourier space. The value of each
point F(x,y) is calculated by multiplying the spatial image
with the corresponding base function and adding the result.
Similarly, the Fourier image can be re-transformed to the
spatial domain. The inverse Fourier transform is exposed in
the equation 2.

N-1 N-1
f ( i,j ) = 1/ N
2
F ( x,y ) e
y2 ( xi / N + yj / N )

(2)
x=0 y=0
Here, normalization term in the inverse
transformation. Sometimes, this normalization is carried out
for the forward transformation instead of the inverse
transformation. To access the result for the above equations, a
double sum has to be obtained for each image point. However,
the Fourier Transform is given by equation 3.
N-1
F( x,y ) = 1/N K ( x, j ) e
y2 pj / N
(3)
j=0
where
N-1
K ( x, j ) = 1 / N f ( i, j ) e
y2 xi / N

i = 0

By using these two equations, initially the spatial domain
image is transformed into an intermediate image using N one-
dimensional Fourier Transforms. This intermediate image is
then transformed into the final image, again use N one-
dimensional Fourier Transforms. Expressing the two-
dimensional Fourier Transform in terms of a series of 2N one-
dimensional transform reduces the number of needed
computationsB. Texture overlapping
Texture overlapping filters decide which parts of the input
image to be patched into the output texture. After finding a
good patch offset between two inputs, the computer is the best
patch seam (the seam yielding the highest possible MRF
likelihood among all possible seams for that offset). The two
overlapped visual cryptic shares images are copied to the
output, cut by max-flow/min-cut algorithm and then stitched
together along optimal seams to generate a new output that is
shown in fig.2. When filtering an overlapped texture, we want
the generated texture to be perceptually similar to the original
image. In this approach, the concept of perceptual similarity
has been formalized by a Markov Random Field (MRF). It
brings an accurate estimation of perceptual effect according to
humans vision.
Figure 2. Texture overlapping method
In most of other techniques, the size of the patch is chosen
a-prior. But this texture overlap filtering technique determine
the optimal patch region for any given offset between the input
and output texture. Finally the performance measure checks
this flexibility for different offsets.
Let us assume a secret image A of N
R
X N
M
. Each pixel of
A can take any one of M different colors or gray-levels. Image
A is represented by an integer matrix A given by equation 4.
A = [a
pq
] N
R
X N
M
(4)
Now M =2 for a binary image, and M =256 for a grayscale
image with one byte per pixel. In a color image, the pixel
value will be an index to a color table, thus M = 256. In a
color image using an RGB model, each pixel has three
integers: R (red), G (green) and B (blue). If each R, G or B
takes value between 0 and 255, we have M = 2563.
The VCS requires taking pseudo-random numbers as input
to guide the choice of the share matrices. Denote the share
matrices in M
p
as S
0
p
,. S
q
Mp-1
, and denote P (S
p
q
) for p =
0, 1 and q = 0, 1,..

Mp-1 as the probability that choosing
the share matrix S
p
q .
Hence the input of the pseudo-random
numbers should guarantee ,that is represented as shown in the
equation 5.
P (S
p
0
) = P (S
p
1
) = .... = P (
Sp
Mp-1
) (5)
In order to choose a share matrix pseudo-randomly in Mp ,
the dealer needs at least log
2
Mp bits pseudo-random
numbers (we will consider the case that log
2
Mp is not an
integer in a later time). Denote B(q) as the binary
representation of integer q with length log
2
Mp, i.e. B(q) is
the binary string that represents q. Without loss of generality,
we assume that when the input pseudo-random number is
B(q), the dealer chooses the share matrix S
p
q
to encrypt the
secret pixel p,and denote P(B(q)) as the probability of
generating the binary string B(q) by the pseudo-random
generator. According to the equation 6,
P(B(0))=P(B(1))= .. =P(B(Mp-1 )) (6)
Vol. 2, No. 11, 2011
74 | P a g e

In fact the cipher texts of the AES and Twofish have
satisfied the above equation, because they have passed the
serial test. Hence, take the AES and Twofish as the pseudo-
random generator.
IV. EXPERIMENTAL RESULTS
In this paper, the experimental simulation is conducted by
using the image processing software package (MATLAB).
The color image (RGB image) is stored in MATLAB as an M-
by- N-by-3 data array that defines red, green, and blue color
components for every individual pixel. The color of each and
every pixel is defined by the combination of the red, green,
and blue intensities stored in each color plane at the pixel's
location

a)Input image b) DFT image
Figure. 3 The experimental result of original input image with out
Using error diffusion and DFT image using Fourier Filtering.

During the experiment, uncompressed image is taken as
input image. Here used (2, 2) VCS scheme and consider the
Lena color image of size 256 X 256 for experimental results
shown in fig. 3(a). This input image is multiplied with the
filter function in a pixel-by-pixel model. To have the resulting
image in the spatial domain, filtered image has to be re-
transformed using the inverse Fourier Transform. The most
simple low pass filter is used to suppress all frequencies
greater than the cut-off frequency and it leaves smaller
frequencies unchanged. In most implementations, cut-off
frequency is taken as a fraction of the highest frequency
represented in the Fourier domain image shown in fig. 3(b).

(a)Share1 image (b) Share2 image
Figure. 4 Experimental result of (2,2 ) Visual cryptography Shares using
error diffusion with the fourier Filtering method.
The (2, 2) VCS scheme is illustrated to introduce the basic
concepts of texture overlapping schemes. In the encryption
process every secret pixel is splitted into two shares. Each
share belongs to the corresponding share image. In the
decryption process the two corresponding shares are joined
together by using OR operation to retrieve the secret pixel.
Two share of a white secret pixel are of the equal while those
of a black secret pixel are complementary as shown in Figure
4(a) and (b).

Figure 5. Decoded image from shares
Error diffusion with texture Overlapping
Consequently a white secret pixel is retrieved by a share
with the combined result of half white sub-pixels and a black
secret pixel is retrieved by all black. Using this basic VCS
Scheme we cant completely retrieve the white Secrete pixel
which generates loss in contrast. In XOR based VCS scheme
where the share images are superimposed using XOR
operation which results in perfect reconstruction of both Black
and white pixels as shown in Figure 5 and sub sampling a 2 X
2 share into a single pixel we get decrypted image of the same
size as original secret image.
The essential parameter indicates the superiority of the
renovation is the Peak Signal-To-Noise Ratio(PSNR). PSNR
is the ratio between the maximum possible power of the signal
and the power of corrupted noise that is articulated in decibels.
Mean Square Error = Error/Size of the image
(7)
The Mean Square Error is the average square of the error
in particular images. The calculation of MSE & PSNR is given
by the equation 8and equation 9.
M N
MSE = 1/ MN [ (Iij I 'ij)2] (8)

i = 1 j = 1

PSNR = 20 * log10 (255 / sqrt(MSE)) (9)

Where, 255 is the maximum possible value of the image.
In general the Peak signal -to-noise ratio for the two shares are
increased and the perceived error for that two shares are
decreased [38]. The imitation result also shows that the
proposed scheme is compared to the existing scheme that is
shown in table 1.
TABLE I. COMPARE THE EXISTING ERROR FILTERING METHOD AND PROPOSED
ERROR FILTERING METHOD.
Error filtering
Method

VC
Scheme
Size of the
Image
Test image - Lena
PSNR in
dB
Error
Ratio
Floyd &
Steinberg Error
Filtering
(Existing)

2-out-
Of-2

256 X 256

11.91

4.74
Discrete Fourier
Filtering
(Proposed)

2-out-
Of-2
256 X 256
36.5826

0.0290

Vol. 2, No. 11, 2011
75 | P a g e
These works are some examples that prove the
improvements and high performance of the color images in
visual cryptography and also reduce the perceived errors.
V. RESULTS AND DISCUSSION
This section provides some experimental results to
exemplify the effectiveness of the proposed method. The
scheme proposed generates meaningful color shares with high
quality as well as the colorful decrypted share by using
Filtering scheme. The performance of the proposed method is
evaluated and exposed in table.1 (that is, our proposed method
is compared with the previous methods). VC can be treated as
a special case in our proposed methods, which means no visual
information is carried by the share. In existing method, shares
carry visual information and there is a tradeoff between the
contrast of the reconstructed image and the contrast of the
share image. This tradeoff is similar to the tradeoff between
the contrast of the reconstructed image and the image quality
of the halftone shares in the proposed methods. Compared
with the existing methods, our method achieves better image
quality, which is given in table2.
Table 2. Reducing the error ratio of the images and Meaningful color shares
with high visual Quality.
Existing Method Proposed Method
No.of
Pixels
Erro
r
Rate
Color
Shares
Err
or
Rat
e
No. of
Pixels
Erro
r
Rate
Colo
r
Shar
es
Error
Rate
1000 11 500 22 1000 8 500 18
2000 19 750 29 2000 9 750 19
3000 28 1000 32 3000 13 1000 20
4000 38 1250 50 4000 19 1250 21
5000 48 1500 60 5000 23 1500 22

Figure 6. Number of pixels Vs Error rate

The results of experiments in which figure 6 and figure 7
indicate that the reducing the error ratio of the images and
meaningful color shares with high visual quality that can
improve the overall performance of the visual cryptography
using texture overlapping and fourier filtering. The error rate
is reduced to 11% compared with the existing scheme.

Fig 7. Color Shares Vs Error rate
VI. CONCLUSION
Some methods for color visual cryptography are not
satisfactory in terms of producing either meaningless shares or
meaningful shares with low visual quality, leading to suspicion
of encryption. In the existing work of color VC the quality of
images being restored depends on error diffusion, other image
degradations due to blurring, transformation and overlapping
were not handled in it.
The color VC focuses on the encryption method, to
produce color Extended Visual Cryptographic System
deploying VIP (Visual Information Pixel) Synchronization and
Error Diffusion for improvement of quality. Error Diffusion
results in the shares with good quality images and VIP
Synchronization regains the actual values before and after
encryption. This paper enhances the image quality on color
visual cryptography using texture overlapping and Fourier
filtering. The proposal in our work improves the image quality
on restored original image from visual cryptic shares by
presenting an efficient color image visual cryptic filtering
scheme. The color image visual cryptic filtering method is
presented here for deblurring effect on the non-uniform
distribution of visual cryptic share pixels.
In the future, color image visual cryptic filtering scheme
proposed in this paper, can be used to maintain digital
document trade marking and licensing with ownership security
schemes. Various multi-party security models used recently
can be adapted in the future for the ownership security.
Privacy preservation techniques (i.e., data transformation and
perturbation) can also be considered for future direction in
providing ownership confidentiality of digital documents.
ACKNOWLEDGEMENT
The authors would like to thank the anonymous reviewers
for their valuable advice and suggestions that have contributed
to the improvement in the quality and presentation of this
paper. In particular, we thank the reviewer who pointed us to
error filtering method in this paper.

0
10
20
30
40
50
0 1000 2000 3000 4000 5000 6000
Number of Pixels
E
r
r
o
r

r
a
t
e

(
%
)
Proposed
Existing
0
10
20
30
40
50
60
70
500 750 1000 1250 1500
Color shares
e
r
r
o
r

r
a
t
e

(
%
)
Proposed
Existing
Vol. 2, No. 11, 2011
76 | P a g e
REFERENCES
[1] InKoo Kang, Member, IEEE, Gonzalo R. Arce, Fellow, IEEE, and
Heung-Kyu Lee, Member, IEEE, Color Extended Visual Cryptography
Using Error Diffusion, IEEE Transactions on image processing, vol. 20,
no. 1, pp. 132-145, January 2011.
[2] M.Naor and A.Shamir, Visual Cryptography , in pro. EUROCRYPT,
1994,pp. 1-12.
[3] M.Naor and A. Shamir, Visual Cryptography, in Advanced in
Cryptology EUROCRYPT94, A. De. Santis, Ed., Vol. 950 of
Lecture Notes in Computer Science, Springer-Verlag, Berlin, PP. 1-
12,1995.
[4] G. Ateniese, C. Blundo, A. D. Santis, and D. R. Stinson, Visual
cryptographyfor general access structures, Inf. Comput., vol. 129, no. 2,
pp. 86106, 1996.
[5] G. Ateniese, C. Blundo, A. D. Santis, and D. R. Stinson, Extended
Schemes for Visual Cryptography, submitted to Discrete Mathematics,
1996.
[6] M. Naor and B. Pinkas, Visual authentication and identification, in
Proc. Advances in Cryptology, 1997, vol. 1294, LNCS, pp. 322336.
[7] E. R. Verheul and H. C. A. Van Tilborg, Constructions and properties
of k out of n visual secret sharing schemes, Designs, Codes,
Cryptog,,vol. 11, no. 2, pp. 179196, 1997.
[8] R. Ito, H. Kuwakado, and H. Tanaka, Image size invariant visual
cryptography, IEICE Trans. Fund. Electron. Commun. Comput. Sci.,
vol.E82-A, no. 10, pp. 21722177, 1999.
[9] C.-N. Yang and C.-S. Laih, Some new types of visual secret sharing
schemes, in Proc. Nat. Computer Symp., 1999, vol. 3, pp. 260268.
[10] L. A. MacPherson, Gray level visual cryptography for general access
structrue, M. Eng. thesis, Univ. Waterloo, Ontario, Canada, 2000.
[11] T. Hofmeister, M. Krause, and H. U. Simon, Contrast-optimal k out of
n secret sharing schemes in visual cryptography, heoreti. Comput.Sci.,
vol. 240, pp. 471485, 2000.
[12] C. N. Yang and C. S. Laih, New colored visual secret sharing
schemes, Designs, Codes Crypt., vol. 20, no. 3, pp. 325336, 2000.
[13] G. Ateniese, C. Blundo, A. Santis, and D. R. Stinson, Extended
capabilities for visual cryptography, ACM Theor. Comput. Sci., vol.
250, pp. 143161, 2001.
[14] G. Ateniese, C. Blundo, A. De Santis, and D. R. Stinson, Extended
Schemes for Visual Cryptography. Theoretical Computer Science, No.
250, pp. 143-161, 2001.
[15] C. Blundo, A. De Bonis and A. De Santis, Improved Schemes for
Visual Cryptography. Designs, Codes, and Cryptography, No. 24, pp.
255278, 2001.
[16] Niranjan Damera-Venkata, and Brian L. Evans, Design and Analysis
of Vector Color Error Diffusion Halftoning SystemsIEEE transactions
on image processing, vol. 10, no. 10, pp. 1552 - 1565 October 2001.
[17] M. Nakajima and Y. Yamaguchi, Extended visual cryptography for
natural images, J. WSCG, vol. 10, no. 2, 2002.
[18] W.-G. Tzeng and C.-M. Hu, Anewapproach for visual cryptography,
Designs, Codes, Cryptog., vol. 27, no. 3, pp. 207227, 2002.
[19] C. C. Lin and W. H. Tsai, Visual cryptography for gray-level images by
dithering techniques, Pattern Recognit. Lett., vol. 24, pp. 349358,
2003.
[20] C. Blundo, P. DArco, A. De Santis, and D. R. Stinson, Contrast
Optimal Threshold Visual Cryptography Schemes , SIAM J. on
Discrete Math. 16, pp. 224-261, 2003.
[21] Y. T. Hsu and L. W. Chang, A new construction algorithm of visual
crytography for gray level images, in Proc. IEEE Int. Symp. Circuits
Syst., 2006, pp. 14301433.
[22] Y. C. Hou, Visual cryptography for color images, Pattern Recognit.,
vol. 36, pp. 16191629, 2003.
[23] C. C. Lin and W. H. Tsai, Visual cryptography for gray-level images
Pattern Recognit. Lett., vol. 25, pp. 349358,2003.
[24] M. Krause and H.-U. Simon, Determining the optimal contrast for
secret sharing schemes in visual cryptography, Combin., Probab.
Comput., vol. 12, no. 3, pp. 285299, 2003.
[25] C.-N. Yang, New visual secret sharing schemes using probabilistic
method, Pattern Recognit. Lett., vol. 25, no. 4, pp. 481494, 2004.
[26] D. Jin, W. Q. Yan, and M. S. Kankanhalli, Progressive color visual
cryptography, J. Electron. Imag., vol. 14, no. 3, p. 033019, 2005.
[27] Ti-Chiun Chang and Jan P. Allebach, Quantization of Accumulated
Diffused Errors in Error Diffusion, IEEE Transactions on image
processing, vol. 14, no. 12, pp. 1960 1975 , December 2005.
[28] C. N. Yang and T. S. Chen, Aspect ratio invariant visual secret sharing
schemes with minimum pixel expansion, Pattern Recogniti. Lett.,
vol.26, pp. 193206, 2005.
[29] W. P. Fang and J. C. Lin, Progressive viewing and sharing of sensitive
images, Pattern Recogniti. Image Anal., vol. 16, no. 4, pp. 632
636,2006.
[30] Z. Zhou, G. R. Arce, and G. D. Crescenzo, Halftone visual
cryptography,IEEE Trans. Image Process., vol. 18, no. 8, pp. 2441
2453, Aug. 2006.
[31] E. Myodo, S. Sakazawa, and Y. Takishima, Visual cryptography based
on void-and-cluster halftoning technique, in Proc. IEEE Int. Conf.
Image Process., 2006, pp. 97100.
[32] G. Horng, T.H. Chen, D.S. Tsai, Cheating in Visual Cryptography,
Designs, Codes and Cryptography, Vol. 38, No.2, pp. 219-236, 2006.
[33] S. J. Shyu, Efficient visual secret image sharing for color images,
Pattern Recognit., vol. 39, no. 5, pp. 866880, 2006.
[34] C.M. Hu and W.G. Tzeng, Cheating Prevention in Visual
Cryptography, IEEE Transactions on Image Processing,Vol. 16, No. 1,
pp. 36-45, 2007.
[35] D.S. Tsai, T.H. Chen, G. Horng, A Cheating Prevention Scheme for
Binary Visual Cryptography with Homogeneous Secret Images, Pattern
Recognition, Vol. 40, No. 8, pp. 2356-2366, 2007.
[36] Zhen He ,Hierarchical Error Diffusion, IEEE Transactions on image
processing, Vol. 18, No. 7, pp. 1524-1534, July 2009
[37] Zhongmin Wang, Gonzalo R. Arce,, and Giovanni Di Crescenzo
,Halftone Visual Cryptography Via Error Diffusion,IEEE Transactions
on information forensics and security, Vol. 4, No. 3, 383-395, September
2009.
[38] Z. M. Wang, G. R. Arce, and G. Di Crescenzo, Halftone visual
cryptography via error diffusion, IEEE Trans. Inf. Forensics Security,
vol. 4, no. 3, pp. 383396, Sep. 2009.
AUTHORS PROFILE
F.R. Shiny malar was born in Nagercoil,Tamil Nadu
State,India in 1986. She studied Information Technology
in theSt.Xaviers Catholic college of Engineering,
Chunkankadai, Kanyakumari District, Tamilnadu
State,India fom 2003 to 2007. She received Bachelor
sdegree from Anna University, chennai 2007. And
received the Master degree from Manonmaniam
Sundaranar University Tirunelveli. currently, she is a
research scholar at the Department of Computer Science and Engineering, in
Noorul Islam Center for Higher Education, Noorul Islam University,
Kumarakoil, Tamilnadu, India; working in the area of image processing under
the supervision of Dr. M. K. Jeya Kumar. She has presented a number of
papers in national conferences and their research interest include image
security , networking and image processing.

M. K. Jeya Kumar received his PhD degree in Mobile
Adhoc Networks from Dr. MGR University,Chennai,
India, in 2010. He is Assistant Professor at the Department
of Computer Application, Noorul Islam University,
Kanyakumari District, Tamilnadu,India. His research
interests include networks andnetwork security, image
processing and softcomputing techniques.
Vol. 2, No. 11, 2011
77 | P a g e
Passwords Selected by Hospital Employees: An
Investigative Study
B. Dawn Medlin
Computer Information Systems
Appalachian State University
Boone, NC, USA
Ken Corley
Computer Information Systems
Appalachian State University
Boone, NC, USA

B. Adriana Romaniello
Economa de la Empresa
Universidad Rey Juan Carlos
Madrid, Spain

Abstract The health care industry has benefitted from its
employees ability to view patient data, but at the same
time this access allows for patients health care records
and information to be easily tampered with or stolen.
Access to and transmission of patient data may improve
care, increase delivery time of services and reduce health
care costs, security of that information may be jeopardized
due to the innocent sharing of personal and non-personal
data with the wrong person. In this study, we surveyed
employees of different size hospitals in various regions of
the state who were willing to share their passwords. Our
findings indicate that employees need further or additional
training in their awareness surrounding password
creation.
Keywords- Passwords; Security; HIPPA; HITECH.
I. INTRODUCTION
Health care records generally include, but are not
necessarily limited to, individual patients health history,
diagnosis, laboratory results, treatments, and the doctors
progress notes. A patients personal information, such as
address, phone number, and social security number, are all
items that may be included and accessible to some or all health
care employees. These records are vulnerable to security
breaches and theft. Both hackers and social engineers have
successfully found ways to penetrate networked health data
systems by simply asking for the information or by finding
weaknesses within the system.
Unfortunately, the largest threat to a health care agencys
security may not be outsiders, but rather their own
employees. Inside employees actually can pose the largest
threat to the security and privacy of information as they can
exploit the trust of their co-workers, and they generally are the
individuals who have or have had authorized access to the
organizations network an d wh o a r e f a mi l i a r with its
internal policies, procedures, and technologies. Additionally,
internal employees can exploit that knowledge to facilitate
attacks and even collude with external attackers (Insider
Threat Research). Due to increased regulations and the
increased opportunities for exploitation that exist in todays
digital world, it is even more important for health
care providers to keep health care records and the
information held within, safe and private. Governmental
agencies have adopted initiatives that specifically address the
issues and rights of health care patients. More specifically, the
security and privacy of health care information is protected by
HITECH (Health Information Technology and Clinical Health
Act) and the Health Insurance Portability and Accountability
Act (HIPAA), requiring health care agencies to do everything
possible to protect their information.
II. BACKGROUND
The electronic accumulation and exchange of p er s on a l
h ea l t h i n f or ma t i on has been promoted as a significant
benefit to health care consumers and providers. Many health
care policy experts believe that broader health information
technology adoption may lead to the availability of more
complete and transparent information, ultimately helping to
contain health care costs while simultaneously improving
health care quality.
Managers must be vigilant in their efforts to protect
patient information as required by several laws. On February
17th, 2009, President Obama signed into law the Health
Information Technology and Clinical Health Act (HITECH) as
part of the American Recovery and Reinvestment Act. The
HITECH Act enhances the security and privacy provisions as
well as the penalties contained in the Health Insurance
Portability and Accountability Act of 1996 (The Health
Information Technology for
Economic and Clinical Health Act (HITECH Act):
implications for the adoption of health information technology,
HIPAA, and privacy and security issues, 2009). This new
law also requires patients be notified in the event of a security
breach.
In addition to HITECH, the basic goal of the Health
Insurance Portability and Accountability Act of 1996
(HIPAA) is to protect the privacy and security of patients and
their medical records. Furthermore, HIPAA addresses
security and privacy measures in relation to passwords, either
directly or indirectly, in the following standards: 1) Security
Management Process [161.308(a)(1)] Health care
organizations must show that they have a consistent set of
internal processes, with implementation that is widespread and
institutionalized. Processes range from establishing criteria for
who has access to what, and who can request certain
resources; to ensuring that access rights are revoked
immediately upon employee termination, 2) Security
Awareness and Training [161.308(a)(5)] HIPAA requires
that staff members be trained and educated concerning the
Vol. 2, No. 11, 2011
78 | P a g e
proper handling of PHI. This basic-level security training
should include measures such as password management, and
3) Access Control [161.312(a)] HIPAA security regulations
require a definition of who has access to PHI within the
organization, as well as the rules determining an individuals
right of access, and the reasons for denying access to some
individuals.
Despite its legal requirements, however, HIPAA standards
are not always followed. As an example, a public posting of
20,000 emergency room patients who had visited Stanford
Hospital in Palo Alto, California, was placed on a commercial
Web site that included the patients names and diagnosis
codes. The hospital confirmed that the i n f o r ma t i o n
remained online for nearly a year (Sack, 2011). Another
example included a laptop that was stolen from a
rehabilitation center containing 660 patients records. The
laptop which was reported stolen from Rancho Los Amigos
National Rehabilitation Center on Feb. 24, 2011 contained at
least 667 patient names, their date of birth and diagnostic
information (Downey, 2011). These are only two examples
of some of the thousands of medical records that are either
stolen or lost each year.
The Federal T r a d e Commission (FTC) and the
Department of Health and Human Services (HSS) i n
2009 i s s ued the first set of HIPAA privacy/security
guidance under the new HITECH Act requirements. The
new g u i da n c e relates to the security br e a ch notification
requirement, that states Under this requirement, health plans
and personal health record (PHR) v e n d o r s mu s t provide
individual notification if there has been a security breach of
protected health information
(http://compliance.utimaco.com/na/tag/hitech- act/).
Additionally, notification must be provided to
individuals in writing within 60 days of discovery of
the breach. If the breach involves more than 500
individuals, notice also must be made in prominent media
outlets and to the Secretary of Housing and Health
Services or to the FTC for PHR vendors (Health IT
Data Breaches: No Harm, No Foul).
For health care administrators, security is enhanced by
using systems tools that are already available, such as Active
Directory and LDAP (Lightweight Directory Access Protocol).
Most likely, one or the other, or a combination of both is
already in use to help in the securing of information. Even
when other front-end access management products, like
IBM Tivoli, Citrix or Sun Microsystems' Java System
Identity Manager are in use, the directory server on the back
end is likely to be Active Directory, LDAP or both.
In addition, more health care agencies may consider
adopting biometrics. Biometrics is the science of identifying
people through physical characteristics. Usually not one
technology but a cluster of several, biometrics uses
fingerprints, handprints, retina scans, voice recognition, facial
structure, and even hand motions while writing a signature-to
identify individuals (Simpson, 2002).
Smart cards may also be used as these operate with a chip
that includes stored memory, and an operating system. A
patient's entire clinical history is stored on the smart card
which can only be accessed via reading devices in a
physician's office, primary care center, hospital, or other
medical institution. Through the use of this device, exposed
paper records will not be a concern. An added benefit of
smart cards is the ability for users to electronically forward
patient information to other health care authorities and
insurers. Specifically, Java-based card technology emerges
as a leading platform because of its ability to support
multiple health care applications securely, while incorporating
biometrics for positive identification and authentication.
A. Issues
Americans hold a strong belief in their right to privacy,
and that belief has been served by the legal system of the
United States. Privacy is also a constitutional concept, as
found in the Fourth Amendment to the U.S. Constitution
(Gostin, 2000). In fact, the preamble to the federal Privacy
Rule, promulgated pursuant to HIPAA, notes that the existence
of a generalized right to privacy as a matter of constitutional
law suggests there are enduing values in American law related
to privacy.
As required by HIPAA as well as other state laws, health
care institutions are required to provide security methods in
order to protect patients information. One such method is
through the authentication of the individual requesting access.
Health care employees are generally subjected to some type of
authentication process. Although there are different ways of
authenticating employees, most systems are based on the use
of a physical token (something one has), secret knowledge
(something one knows) or biometrics (something one is)
(Burnett & Kleiman, 2006).
In todays health care institutions, the most common
authentication mechanism is still the simple use of a password
(something one knows or creates). This type of authentication
method can offer to employees the ability to quickly enter into
a system, but human practices such as using the same
password on different systems and writing down a password
may degrade the quality of password security (Pfleeger and
Pfleeger, 2007).
The authentication method of individuals creating their
own passwords is not atypical. For health care organizations
the password functions like the key t o a l o c k , anyone who
has it can get in to see the patients information. Toward that
end, there have been recommendations from governmental
agencies to hospitals on how to construct a password. One
of the first guidelines in creating good passwords was
published in 1985 by the Department of Defense and is still
relevant today (Department of Defense. 1985). Their
guidelines recommended the following: 1) passwords must be
memorized; 2) passwords must be at least six characters
long, 3) passwords must be replaced periodically, and 4)
passwords must contain a mixture of letters (both upper-
and lowercase), numbers, and punctuation characters.
Most networks administrators and security experts would
concur with all of the above Department of Defense
recommendations, however, that was in 1985 when the advice
was given and when social engineering as well as other types
Vol. 2, No. 11, 2011
79 | P a g e
of attacks were not as common as they are today.
According to CERT (the Computer Emergency Response
Team), the advice to use upper and lower case alpha
characters for Novell and/or VMS systems is useless since
both of these systems are case insensitive.
B. Problems
Many of the deficiencies of password authentication
systems arise from the limitations of human cognitive ability
(Pond et al., 2000). If humans were not required to remember
a password, a maximally secure password would be one
with maximum length that could consist of a string of
numbers, character, and symbols. In fact, the requirements to
remember long and complicated passwords are contrary to
the way the human memory functions. First, the capacity of
human memory in its capacity to remember a sequence of
items is temporally limited, with a short-term capacity of
around seven items plus or minus two (Kanaley, R., 2001).
Second, when humans remember a sequence of items,
those items cannot be drawn from an arbitrary and
unfamiliar rang, but must be familiar 'chunks' such as
words familiar symbols. Third, the human memory thrives on
redundancy.
In fact, studies have shown that individuals short term
memory will retain a password for approximately 30 seconds
thereby requiring individuals to attempt to immediately
memorize their passwords. It has also been shown that if an
individual is interrupted before they fully memorize the
password; it will fall out of their working memory and most
likely be lost.
Also, if an individual is in a hurry when the system
demands a new password, individuals must sacrifice either the
concentration of the critical task at hand or the recollection
of the new password. Related to this issue is having to create
the content for this new quickly demanded password. The
pressure to choose creative and secure passwords quickly
generally results in individuals failing in their attempt to
memorize this new password. For health care organizations
this can result in reset rates at one per reset per every four to
five users per month (Brostoff and Sasse, 2001).
In order to combat the issue of having to remember so
many different passwords some users have resorted to the
selecting familiar terms such as a pet or family name, their
own name, their phone number, or other common terms that
could be found in a dictionary. British psychologist
Helen Petrie, Ph.D., a professor of human/computer
interaction at City University in London analyzed the
passwords of 1,200 British office workers who participated in
a survey funded by Ce n t r a l Ni c , an Internet domain-name
company in 2001. She found that most individuals'
passwords fell into one of four distinct password categories
which were family, fan, fantasists, and cryptic.

The first category of family," comprised nearly half of
the respondents. These individuals selected their own name,
the name of a child, partner or pet, birth date, or significant
number such as a social security number. Further, Dr.
Petrie found that individuals also choose passwords that
symbolized people or events with emotional value or ties.
One third of the survey participants were identified as
"fans," using the names of athletes, singers, movie stars,
fictional characters, or sports teams. Dr. Petrie al so
found that these individuals wanted to align themselves with
the lifestyle represented by or surrounded around a celebrity
status. Two of the most popular names were Madonna and
Homer Simpson.
Fantasists made up eleven percent of survey responses. Dr.
Petrie found that their passwords were comprised of sexual
terms or topics. Some examples included in this category
were terms such as sexy," "stud and "goddess."
The final ten percent of p a r t i c i p a n t s were identified
as "cryptics. These users were seemingly the most security-
conscious, but it should also be noted that they were also the
smallest of all of the four identified categories. These
individuals selected unintelligible passwords that included a
random string of letters, numerals, and symbols such as
Jxa+157.
Self-created computer passwords are generally personal,
and they reflect the personalities of millions of people as they
attempt to summarize their life through a few taps on the
keyboard. As psychologists know, people and personalities
are often very predictable in the aggregate (Andrews, 2004),
as may be their choices of passwords. Psychologists have
found that humans can store only five to nine random bits of
information in their short-term memory, making it difficult to
remember long and complicated passwords. Therefore, users
have often chosen passwords with personal meanings that they
can associate with something in their long-term memory.
III. RESEARCH METHODOLOGY
A. Data Collection
To obtain a fair statistical representation of the password
security used in relation to health care organizations, a survey
was given to employees at five hospitals of various sizes and
in different regions of the state. Hospital administration
approval was obtained, but the administration did not endorse
the survey to respondents, nor did they ask them to participate.
The data set was comprised of 118 responses. Data was
gathered to not only determine how many employees would
disclose their passwords and other personal information such
as their address, phone number and email, but also simulated
the types of information individuals were willing to share with
co-workers, colleagues, or friends of colleagues. The
information that employees were willing to share, including
their passwords and other personal information, would
certainly make it easier to hack into a system instead of having
to guess at the necessary authentication information.
B. Analysis and Results
As seen in Table 1 (labeled password categories), half of
the respondents created passwords consisting of family
names, including their own name or nickname, the name
of a child, or significant other. Interestingly, the findings
noted in Table 2 indicate that most r esponden t s wer e
often required to use a password to access systems, but
Vol. 2, No. 11, 2011
80 | P a g e
rarely changed their passwords. As further indicated, most
of the respondents used the same password on multiple
accounts. The practice of rarely c h a n g i n g passwords
and/or using the same password for mu l t i p l e accounts
wo u l d assist social engineers, thus allowing them to easily
attain access to one system and possibly more.
TABLE 1. PASSWORD CATEGORIES

Variable
Name
Question Answers N Mean Std Dev

Family
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.50

0.50

Cryptic
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.05

0.22

Number
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.45

0.50

Fan
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.15

0.95

Faith
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.03

0.18

School
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.02

0.13

Fantasy
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.00

0.00

Place
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.14

0.34

Other
Does your
password
fit into this
category?
1 = Yes, 0 = No

118

0.51

0.50
Additionally, as seen above in Table 1, the largest
category of password choices included some type of
relationship to a family name being reported at fifty percent
(50%). Fifteen percent (15%) of the respondents self-reported
the inclusion of fan-based words, which could include
names of athletes, singers, movie stars, and fictional characters
or sports teams. Place was the next highest category, with
fourteen percent (14%), using another identifiable piece of
information such as the city where the employee works/lives.
The smallest of all of the self-identified password
categories was fantasy, followed closely by the categories
of school and faith. Five percent (5%) of the employees
selected the cryptic category, suggesting that these
employees are security-conscious since that category includes
passwords that are unintelligible.

TABLE 2. PASSWORD STATISTICS

As noted in Table 3, in the area of password and security
training, most of the respondents, fifty-four percent (54%)
indicated that their employer had offered password security
training, with fifty-eight percent (58%) of the hospitals
offering some other type of security awareness training (Table
3). Attendance by the employee in either a current password
or security awareness training program was measured on a
Likert scale of 1 being last week and 5 being never. The
employees indicated that currently, they almost never attended
the security awareness programs.
TABLE 3: PASSWORD TRAINING
IV. DISCUSSION
This study reveals several interesting findings. As noted
earlier, most employees used the same passwords on multiple
accounts, even though they frequently changed them. The
actions of repeatedly using the same password are contrary to
Variable
Name
Question Answers N Mean
Std
Dev

Pass_Freq
How often do use
a
password to
access systems?

1= Very Often
5= Never

118

1.23

0.59

Pass_
Change
How often do
you change
your
passwords?
1= Very Often
5= Never

117

2.85

1.13

Reuse

Most people
use the same
password on
multiple
accounts. How
often do you do
this?

1= Very Often
5= Never

118

2.47

1.32
VAR NAME QUESTION ANSWER NO MEAN
STD
DEV
Pass_Train
Does your
employer offer
password
security
training?
1 = Yes,
0 = No
115 0.54 0.50
Awar_Train
Does your
employer offer
any other
security
awareness
training?
1 = Yes,
0 = No
113 0.58 0.50
Current_Train
When was the
last time you
participated in
either a
password or
another
security
awareness
training
program?
1= Last
week,
5 = Never
115 4.09 1.08
Vol. 2, No. 11, 2011
81 | P a g e
suggested recommendations by most security experts, because
a hacker who gained access to one account could more easily
access other systems. Requiring individuals to maintain a new
password for each system or application would obviously
make systems more secure but is in conflict with humans
short-term human memory capabilities. Employees may
consider it necessary to include familiar names, places, and
numbers in their passwords so that they can easily recall them.
Though most employees indicated that their employers
offered password security training either very often or often, it
appears that either the types of training are not very effective
or that the employees did not take it very seriously.
V. CONCLUSION
Findings of the present study indicate that employees are
willing to share personal information with co-workers and
friends of co-workers. Seventy-three percent (73%) of the
employees shared information that a social engineer could use
to create a profile of an employee and gain access to the
employers network and other confidential patient
information. It is imperative that employees understand the
consequences of sharing information as well as the importance
of creating and maintaining strong passwords.
The simulation that was carried out during this study
demonstrated that many employees may currently be in
violation of HIPAA and HITECH regulations due to their
willingness to share their information and their practice of
creating weak passwords, thus allowing for easy access into a
system. Hospitals and other health care agencies must identify
ways to educate employees regarding HIPAA and HITECH
regulations to protect patients and practices to create a long
password, but on the other hand offers it freely to others. This
study demonstrated that many employees may currently be in
violation of HIPPA and HITECH regulations due to their
willingness to create weak passwords and to share them with
strangers through our survey instrument.
REFERENCES
[1] 6 Password Protection Tips That Every Computer User MUST Know!.
Retrieved on October 23, 2011 from
http://www.richtechgroup.com/business/2011/04/6-password-protection-
tips-that-every-computer-user-must-know/
[2] Andrews, L.W. (2004). Passwords reveal your personality. Retrieved
March 13, 2007, from http://cms.psychologytoday.com/articles/pto-
20020101-000006.html.
[3] Brostoff, S., & Sasse, M. A. (2001). Safe and Sound: a safety-critical
approach to security. Position paper presented at the New Security
Paradigms Workshop 2001, Cloudcroft, New Mexico.
[4] Burnett, M. & Kleiman, D. (2006). Perfect Passwords. Selection,
Protection, Authentication. Syngress.
[5] Data Breach Harm Analysis from ID Analytics Uncovers New Patterns of
Misuse Arising from Breaches of Identity Data. Retrieved on November
12, 2009, http://www.idanalytics.com/news_and_events/20071107.html
[6] Department of Defense. (1985). Password Management Guideline.
http://www.alw.nih.gov/Security/FIRST/papers/password/dodpwman.txt
[7] Downey (2011). Laptop Stolen from Rehab Center with Over 660 Patient
Records. KTLA News. Retrieved on October 23, 2011 from
http://www.ktla.com/news/landing/ktla-laptop-stolen-from-downey-
rehab,0,3270396.story
[8] Georgetown University Information Security. Retrieved November 12,
2009, from
[9] Gostin, L.O. (2000). Public Health Law: Power, Duty, Restraint.
University of California Press , Berkeley, CA. pp. 132-134.
[10] Gragg, D. (2007). A Multi-Level Defense Against Social Engineering.
SANS.
http://www.sans.org/reading_room/whitepapers/engineering/920.php.
Human Memory. Intelegen, Inc. Retrieved on December 1, 2009 from
http://brain.web-us.com/memory/human_memory.htm.
[11] Health Law Alert. Retrieved on November 15, 2009 from
http://www.nixonpeabody.com/publications_detail3.asp?ID=2621.
[12] Health IT Data Breaches: No Harm, No Foul. Retrieved on Novemer 12,
2009 from http://compliance.utimaco.com/na/tag/hitech-act.
[13] Hupp, M. (2007). Protecting patient medical records from the nosy.
Retrieved on November 30, 2009 from
http://www.bizjournals.com/milwaukee/stories/2007/11/12/focus3.html?
t=printable.
[14] Insider Threat Research. Retrieved December 1, 2009 from
http://www.cert.org/insider_threat.
[15] Internet Identity Theft And Password Security Tips. Retrieved on
October 23, 2011 from http://www.combat-identity-theft.com/internet-
identity-theft.html
[16] Kanaley, R. (2001). Login error trouble keeping track of all your sign-
ons? Heres a place to keep your electronic keys, but you better
remember the password. San Jose Mercury News, 3G.
[17] Pfleeger, C.P. & Pfleeger, S.L. (2007). Security in Computing. Fourth
edition. Prentice Hall.
[18] Pond, R., Podd, J., Bunnell, J., Henderson, R. (2000). "Word Association
Computer Passwords: The Effect of Formulation Techniques on Recall
and Guessing Rates," Computers & Security, 19, 645-656.
[19] PROTECTING HEALTH INFORMATION. RETRIEVED ON NOVEMBER 12,
2009 FROM HTTP://WWW.HHS.GOV/NEWS/FACTS/PRIVACY.HTML.
[20] Sack, K. (2011). Patient Data Posted Online in Major Breach of Privacy.
New York Times. Retrieved on October 23, 2011
http://www.nytimes.com/2011/09/09/us/09breach.html?pagewanted=all)
[21] Simpson, R.L. (2002). Nursing Management. Chicago: 33(12), 46-48.
[22] The Health Information Technology for Economic and Clinical Health
Act (2009). Retrieved on October 23, 2011 from
http://www.nixonpeabody.com/publications_detail3.asp?ID=2621
[23] Thompson, S. T. Helping the Hacker? (2006) Library Information,
Security, and Social Engineering. Information Technology and Libraries.
Chicago: 25(4), 222-226.
AUTHORS PROFILE
B. Dawn Medlin is the Chair and Professor in the Department of Computer
Information Systems and the Co-Director of the Center for Advanced
Research on Emerging Technologies, John A. Walker College of Business, at
Appalachian State University in Boone, NC. During her 24 years of teaching
she has taught courses such as Web 2.0 Technologies in Business,
Introduction to Gaming, Advanced Security, Issues in E-Commerce, She has
published in journals such as The Journal of Information Systems Security,
Information Systems Security, International Journal of Electronic Marketing
and Retailing, and the International Journal of Healthcare Information
Systems and Informatics. Additionally, she has taught at the Universit
d'Angers and Addis Ababa University in Ethiopia.

Ken Corley is an Assistant Professor of Computer Information Systems
at Appalachian State University. He received his Ph.D. from
Auburn University in Auburn, Alabama, USA. His current research
interests include information privacy & security, computer and human
interaction, and sustainability.

B. Adriana Romaniello, originally from Uruguay, is an Interim Associate
Professor of Management at the Department of Business Administration at
University Rey Juan Carlos (Madrid, Spain) who came from University Carlos
III (Madrid) where she teaches Corporate finance. She holds a Ph. D. in
Business Administration at University Complutense of Madrid and a
Licentiate degree in economics and business from University of La Republica
(Uruguay). She teaches management, organization theory and organizational
design to undergrads and doctoral students. Adriana's research interests focus
on information security, organization theory and competitive strategy.
Vol. 2, No. 11, Nov 2011
82 | P a g e
Current Trends in Group Key Management

R. Siva Ranjani
1

Research Scholar, Dept.of CS&SE
Andhra University, Visakhapatnam,
Andhra Pradesh, India,

Dr.D.Lalitha Bhaskari
2

Associate Professor, Dept.of CS&SE
Dr.P.S.Avadhani
3

Professor, Dept.of CS&SE

AbstractVarious network applications require sending data
onto one or many members, maintaining security in the large
groups is one of the major obstacles for controlling access.
Unfortunately, IP multicast is not providing any security over the
group communication. Group key management is a fundamental
mechanism for secured multicast. This paper presents relevant
group key management protocols. Then, we compared them
against some pertinent performance criteria. Finally, we discuss
the new research directions in group key management.
Keywords-Multicast; group key management; security member
driven; time driven.
I. INTRODUCTION
With rapid growth in the internet, people using the group
communication in applications such as paying TV,
transmission of video and audio, updating software, military
applications, video games etc. In recent decades, the focus is
mainly on the security issues involved in the group
communication. When the group uses the unicast
communication, one sender is sending the data stream onto one
group member. In multicasting, the group member is sending
the data onto other group members. Security is main focused
area in group communication. Group key management is the
fundamental mechanism provides the security in group
communication. In this, security is achieved by sharing a
common key among the group members. The message packets,
those are going to transmit should be encrypted with the shared
key.
Group key management is mainly focusing on the key
generation and distribution of key among the group members.
All the group members should participate in the secure
distribution, creation and revocation of the keys [1]. The
communication session in group key management is managed
by two entities: Group Controller (GC), responsible for key
generation, distribution and rekeying for membership change
and Key Server (KS), responsible for maintaining the keys
and distributing the keys.
The scenario of group communication is shown in the
figure 1. Each member in the group is having two keys (TEK
and KEK). The TEK (Traffic Encryption Key) is used for
encrypt, decrypt and authenticate the data transfer. The TEK
for the group member is generated by the local manager. The
KEK (Key Encryption Key) is used for encrypt the TEK. To
multicast the message (m) secretly the sender encrypts the
message with TEK using the symmetric key algorithm. At the
receiving side, the receiver decrypts the message (m) with
TEK. In the group communication, the members in the group
are not fixed, members can join / members can leave the group.
So we need to secure the sending message to be received by the
group members at that instance. When member is leaving, the
KS must generate a new TEK and distribute the key secretly to
all other members except leaving one. This process is known as
rekeying. From the figure, we observe that Key Server is
sharing a secret key called Key Encryption Key (KEKi) with
each group member. When the member is leaving, the KS
generates a new TEK : TEK1, encrypted with their KEKi and
sends it to all other group members except leaving one. So the
leaving member does not know the new TEK1, to decrypt the
future messages shared in the group.
When a new member is joined in the group, first it must be
authenticated by the GC. After that, the KS checks the rights of
the new member and adds the member in the future message
transformation session. The KS generates a new secret KEKj
and shared with the new member mj. In order to restrict the
new member form past data access the KS generate the new
TEK : TEK1, encrypted with KEKi and then sends to all the
group members along with the new joined one.
II. GROUP KEY MANAGEMENT PROTOCOL
As defined by Menezes et al. in [2], Group Key
Management is the set of techniques and procedures used for
the establishment and maintenance of keys among members to
form the group. According to Hutchison [3], group key
management can be classified into three categories.
Centralized Group Key Management ProtocolsKey
distribution is achieved by a single entity i.e Key Distribution
Center (KDC), also known as Central Authority (CA). The
Central Authority maintains the entire group, allocates the
individual KEK to group members. It is also responsible for
sharing the common TEK among all the group members.
Decentralized Group Key Management Protocols In
Decentralized Group, the group is splitting into several
subgroups. Each subgroup is managed by subgroup controller.
In this approach, the hierarchy of sub group controllers shares
the labor in transferring TEK to group members. This
management will reduce the load on the KDC.
Distributed Group Key Management Protocols In
Distributed Group key management either all the group
members or only one member is involved in group key
generation. No group controller is present; this will improves
the reliability of the overall system.
Vol. 2, No. 11, Nov 2011
83 | P a g e

Figure 1: Group with KS and 5 Group Members
A. Centralized GKMP
Assume there is a group with n members. The Key server
(KS) is the centralized group manager which stores information
about all the group members. The KS takes n Key Encryption
Keys (KEK) and each of them is shared with one member. The
KEK is the secret key used for encrypting the group key
(TEK). In the figure, the centralized group key protocols again
sub divided into three categories depending on the technique
used in distributing the TEK among group members.
a) Pair wise Keys:
In this sub type the KS shares a secret key called Key
Encryption Key (KEK) with all the group members. This key is
used for establishing the secured channel between the KS and
the member for transferring the TEK securely whenever the
key is required.
Harney and Muckenhirn [4],[5] proposed Group Key
Management Protocol (GKMP), in this KS shares a secret key
(KEK) individually with each active member and KS generates
a Group Key Packet (GKP) that contains a Group TEK
(GTEK) and a group KEK (GKEK). Chu et. al Protocol is
proposed by Chu et. al [6], a Group leader shares a secret Key
Encryption Key (KEK) with each group member.
b) Broadcast Secrets:
In this protocol, the KS broadcasts the rekey information to
all the group members. Chiou and Chen [9] proposed Secure
Lock protocol, in this key server uses a single broadcast to
establish the group key or for sending rekey to the entire group
in the case of join or leave membership.
c) Keys Hierarchy:
In pair wise key approach, the KS establishes individual
secure channel with the members and uses this channel for
sending the TEK updates. This mechanism increases the update
message overhead. In order to reduce message overhead, the
Key Server in this approach, shares a secret key with subgroup
in addition to individual channel. When the member leaves the
group the KS uses the sub group secret key to distribute the
new TEK, which are not known by the leaved member.
Nemaney et al.[26] proposed hierarchical group key
management that increases the efficiency of the key. Following
section describes some of the protocols using this re-key
mechanism.
Wong et.al and Wallner et. al[11][14] proposed Logical
Key Hierarchy (LKH) protocol. In this protocol, KS is the root
of the tree and maintains a tree of keys. In this protocol, each
node stores at most 1+log2(n) rekey messages.
McGraw and Sherman proposed One-Way Function Tree
(OFT) protocol and this is an improvement over the LKH.
Here, nodes KEK is calculated by the member rather than
attributed by the KS. Each node in this protocol is maintaining
its blended sibling keys and its leaf secret key also maintains
the blinded secret KEKs of its ancestors.
Canetti et. al proposed One-Way Function Chain Tree
(OFCT) protocol [13]. This protocol works similar to OFT but,
a pseudo random generator is used to generate the new KEKs
rather than a one-way function, and it is done only during user
removal.
Efficient Large Key distribution (ELK) [15] approach uses
the pseudo random function for generating the new KEK when
a membership change takes place. Waldvogel et.al [16]
proposed Centralized Flat Table Key Management (CFKM)
protocol, this approach, uses the flat table concept in order to
reduce the number of keys maintained by the KS. Flat table
consists of one TEK and 2w entries for KEKs, where w is the
number of bits in identifier of a member. Wong et.al protocol is
the extension of the LKH protocol [14]. The LKH uses the
binary tree for key distribution; wong et.al uses the k-ary tree.
Comparison of centralized group key management
protocols:
Table I compares the centralized group key management
protocols. The efficiency of the protocol can be compared
against the following criteria: 1 affect n, forward and backward
secrecy, storage requirements at KS and group member,
collusion, join re-key overhead and leave rekey overhead.
a) Decentralized Group key Management Protocols
The group members are arranged into some subgroups, and
each subgroup has a controller called key manager. The key
managers of the subgroup share the labor of distributing the
TEK to group members in order to avoid bottle necks and
single point of failure. Decentralized group key management is
categorized into member ship driven and time driven re-keying.
Ballardies Scalable Multicast Key Distribution (SMKD)
protocol [18] propose a group key distribution method based on
the Core Based Tree (CBT) multicast routing protocol. In CBT
architecture, the multicast is rooted at main core. In Intra
Domain Group Key Management (IGKMP) [17], the network
divides into administratively scoped areas. This protocol is
having Domain Key Distributor (DKD) and Area Key
Distributor (AKD).
Source
TEK
M1
TEK
KEK1
M2
M4
GC/KS
M5
TEK
KEK5
M3
TEK
KEK3
TEK
KEK1
KEK3
KEK5
-------- Multicast
M1, M3, M5 group
Vol. 2, No. 11, Nov 2011
84 | P a g e
TABLE I : COMPARISON OF CENTRALIZED GROUP KEY
MANAGEMENT
Protocol
Type
Server
Storage
Rekeying overhead
Member Join Member leave
Poovendran et al n+2 2 2n
Dunigan & chao n+2 2 2n
Chu et. Al
n+2 2 2n
Secure Lock
2n 2 0
LKH
2n-1 log2(n)+1 2log2(n)
OFT
2n-1 log2(n)+1 log2(n)+1
OFCT
2n-1 log2(n)+1 log2(n)+1
ELK
2n-1 log2(n)+1 log2(n)+1
CFKM
2I+1 2I 2I
Where n: number of group members I: number of bits in a number id
The DKD is responsible for generating group TEK and is
propagated to the group members through AKD. The DKD and
AKDs belong to multicast group called All-KD-Group.
In Hydra protocol [19] the group is organized into sub
groups. Each sub group has a server called Hydra Server (His)
responsible for controlling the sub group. BAAL protocol has
three entities: First is the Group controller (GC), responsible
for maintaining the participant List (PL) and creating and
sending the group key TEK to member through local
controller. Second is Local Controller (LC), responsible for
managing the keys in subnet, receives the new TEK and
distributes to members connected to subnet. Third is Group
member. IOLUS protocol is the frame work of a hierarchy of
multicast subgroups to constitute virtual group [20]. Each
subgroup is managed by a Group Security Agent (GSA),
responsible for managing key inside the sub group.
Cipher Sequences is a proposed framework for multicast
security [21], based on Reversible cipher sequence. The
multicast tree is rooted at source and the leaves are group
members. Challel et.al proposed Scalable Adaptive Key
Management Scheme (SAKM) protocol. This protocol tackles
the scalability issue. SAKM tackles the scalability by
organizing the group into clusters.
b) Time Driven Approach
In time driven approach, the TEK is changed after specified
amount of time. When the member leaves or joins in the group
they will not excluded or appointed immediately, need to wait
for the beginning of the new interval of time.
Briscoe proposed MARKS protocol, suggests a slicing the
time length into small portions of time and uses a different key
for encrypting each slice. The encryption keys occupied at
leaves in BST that is generated from a single seed.
Setia et al [22] describe a scalable approach based on time-
driven called Kronos. In this protocol, Setia denotes the group
with a birth and death process model and discussed the model
in two occasions: correlation subscriber behavior and
independent subscriber. The operation of Kronos is similar to
that of IGKMP. In Dual Encryption Protocol (DEP), the group
is divided hierarchically into sub groups and the sub group is
managed by sub-group manager (SGM). In YANG et. al
Protocol [23] approach the multicast group is organized into a
set of sub groups, KS manages each subgroup. The KS is
responsible for rekeying the members in the subgroup
periodically. Scalable Infrastructure For Multicast Key
Management (SIM-KM) uses the proxy encryptions. SIM-KM
uses the proxy function that converts the cipher text for one key
into the cipher text for another key.
c) Comparison of Decentralized Group Key
Management Protocols
In this different Group controllers are used to manage the
subgroups. Table II compares the decentralized group key
management protocols. Attributes that are used for evaluating
the performance of decentralized protocols are key
independence, decentralized controller, local rekey, key-data
transformation, rekey per membership and type of
communication.
TABLE II: COMPARISON OF DECENTRALIZED GROUP KEY
MANAGEMENT
Protocol Key
Inde
pen
dent
Decentralized
Controller
Loc
al
re-
key
Key
Vs
Data
Re-
key
Comm
unicati
on
Type
Mana
geme
nt
Key
Serv
er
SMKD Yes Yes Yes No Yes No Both
IGKMP Yes Yes Yes No Yes Yes Both
Hydra Yes Yes Yes No Yes Yes Both
Baal Yes Yes Yes No Yes No Both
MARKS No Yes - No Yes No Both
Kronos No Yes Yes No Yes No Both
DEP Yes No No No Yes No Both
Iolus Yes Yes Yes Yes Yes Yes 1 to n
KHP Yes Yes No Yes No Yes Both
Cipher
Sequences
Yes No No No Yes Yes 1 to n
SAKM Yes Yes Yes Yes Yes No 1 to n
Yang et al Yes Yes Yes Yes Yes Yes 1 to n
SIM-KM No No No Yes Yes Yes 1 to n
Both: 1 to n and n to n
d) Distributed Key agreement Protocols
In distributed key agreement protocol, the group members
are participated in the establishment of a group key, further
classified into three categories: Ring-based, hierarch based and
broadcast based.
e) Ring Based
In this, cooperation of group members forms a virtual ring.
In Ingemarson Et Al. protocol, all the group members are
organized into a virtual ring and the Group Diffie-Hellman
(GDH) protocol uses the extension of Diffie Hellman algorithm
for group key generation.
f) Hierarchy based cooperation
The group members are arranged in a tree hierarchy for
group key generation. In OCTOPUS, the entire group is
divided into four sub groups. The leader member in the
subgroup is responsible for collecting the intermediate
subgroup values and calculates the intermediary DH value.
Steer et. al proposed Skinny Tree (STR) protocol, uses the tree
structure. The leaves associated in the tree are group members;
each leaf is identified by its position. Diffie-Hellman Logical
Vol. 2, No. 11, Nov 2011
85 | P a g e
Key Hierarchy (DH-LKH), proposed by Perrig et al.[24] is
variant of STR and uses binary tree. The binary tree built from
bottom to top. Distributed Logical Key Hierarchy (D-LKH)
protocol uses the notion of sub-trees, agreeing on a mutual key.
Distributed One-way Function Tree (D-OFT) approach using
logical key hierarchy in a distributed fashion was proposed by
Dondeti et al, uses the one-way function tree proposed by
McGrew and Sherman. Every group member is trusted with
access control and key generation. In Fiat and Naor protocol,
each member broadcast a single message to other participants
in order to agree on a common secret.
g) Broadcast based approach
In this approach, group key is generated by broadcasting
the secret messages and distributing the computations among
the group members. Burmester And Desmedt Protocol is a
three round protocol with member generation, broadcasting and
group key computations.
Boyd proposed Conference Key Agreement (CKA)
protocol, where all the group members contributed to generate
the group key.
h) Comparison of Distributed Group Key Management
Protocols:
All the members in the group are involved in the
computation of group key or generated by one member in the
group. Table III compares the distributed group key
management protocols. Attributes to evaluate the efficiency of
distributed key management protocols are a number of rounds,
number of messages, DH key and leader requirement.
TABLE III: COMPARISON OF DISTRIBUTED GROUP KEY
MANAGEMENT
Protocol No of
Rounds
No of Messages DH
Key
Leader
Req. Uni-cast Multi-cast
GDH n n-1 n Yes No
Ingemarson et. al n-1 n(n-1) 0 Yes No
Octopus 2(n-1)/4
2
3n-4 0 Yes Yes
STR n 0 N Yes No
DH-LKH log2(n) 0 log2(n) Yes No
D-LKH 3 N 1 No Yes
D-OFT log2(n) 2log2(n) 0 No No
D-CFKM n 2n-1 0 No No
Fiat et. Al 2 N N Yes Yes
Bermester et 3 0 2n No No
CKA 3 n-1 N No Yes
III. CURRENT RESEARCH DIRECTIONS
A group Key Management application in mobile networks,
ad hoc networks, e learning, and peer-to-peer networks is
prevalent. Many new protocols are proposed as existing key
management protocols are no more suitable for these areas.
Jiang and Hu [8] classified current group key management
protocols as stateless, self-healing, distributive, reliable,
adaptive and mobile-based. Among these protocols reliability
and distributiveness are by default provided by the group key
management protocols. Scalability of stateless group key
management protocols is enhanced by reducing the degree of
the polynomial functions with the help of the decentralized
subgroup managers. Junbeom and Hyunsoo [12] proposed a
decentralized multi-group key management scheme for
stateless group members. Self-healing and rekeying are
becoming target areas in the group key management protocols.
Key Server transmits group key updating messages when there
are some changes in membership states. A self-healing protocol
can recover certain number of existing and/or future group
keys. First, self-healing key distribution protocol [8] was
proposed by J.Staddon et.al [25] which is based on polynomials
and Angelo [10] provided an efficient self-healing scheme for
LKH. Challal et al.[7] proposed adaptive group key
management protocol and there is need of extensive research
should be done in this
IV. CONCLUSION
This paper focused on group key management, secured
distribution of session keys and refreshment of the keying
material. Reviewed so many group key management protocols
and placing them into three main classes: centralized,
decentralized and distributed protocols, which try to minimize
the requirements of KDC and group members. Centralized key
management is easy to implement but more overhead on single
member. The decentralized key management follows the
hierarchical sub grouping and it is harder to implement.
Distributed key management is simply not scalable. From the
comparison tables, we analyze that no unique solution that can
achieve all the requirements. Hence, it is important to
understand fully the requirements of the application before
selecting a security solution. A solution for secure group
communication should complement a multicast application
rather than drive its implementation. The usage of security
mechanism for secure group communication should be made
transparent to the user and it should also work well with other
protocols.
REFERENCES
[1] David Manz, Jim Alves-Foss and Shanyu Zheng, Network Simulation
of Group Key Management Protocols, Journal of Information
Assurance and Security, pp. 67-79, January 2008.
[2] A. J. Menezes, P. C. van Oorschot, and S. A. Vanstone, Handbook of
Applied Cryptography. CRC Press, 1996.
[3] S. Rafaeli and D. Hutchison, A Survey of Key Management for Secure
Group Communication, ACM Comput. Surv., vol. 35, no. 3, pp. 309
329, 2003.
[4] H. Harney and C. Muckenhirn. Group Key Managemant
Protocol(GKMP) Architecture. July 1997, RFC 2093.
[5] H. Harney and C. Muckenhirn. Group Key Managemant
Protocol(GKMP) Specification. July 1997, RFC 2094.
[6] H.H. Chu, L. Qiao, and K. Nahrstedt. A Secure Multicast Protocol with
Copyright Protection. ACM SIGCOMM Computer Communications
Review, 32(2):42:60, April 2002.
[7] Y. Challal, H. Bettahar, and A. Bouabdallah. SAKM: A Scalable and
Adaptive Key Management Approach for Multicast Communications.
ACM SIGCOMM Computer Communications Review, 34(2):5570,
April 2004.
[8] Bibo Jiang, Xiulin Hu, A Survey of Group Key Management,
International Conference on Computer Science and Software
Engineering, 2008, IEEE, DOI 10.1109/CSSE.2008.1282
[9] G. H. Chiou and W. T. Chen. Secure Broadcast using Secure Lock.
IEEE Transactions on Software Engineering, 15(8):929934, August
1989.
[10] Angelo Rossi, Samuel Pierre and Suresh Krishnan, An Efficient and
Secure Self-Healing Scheme for LKH, Journal of Network and Systems
Management, Vol. 18, Number 3, 327-347
Vol. 2, No. 11, Nov 2011
86 | P a g e
[11] Debby M. Wallner, Eric J. Harder, and Ryan C. Agee. Key management
for multicast: Issues and architectures. Internet draft, Network working
group,september 1998, 1998.
[12] Junbeom and Hyunsoo, 2009, A decentralized multi-group key
management scheme, IEICE Transactions Communications, Vol E-92-
B, No.2, Feb 2009
[13] R. Canetti, T. Malkin, and K. Nissim, Efficient Communication-
Storage Tradeoffs for Multicast Encryption, in EUROCRYPT. New
York, NY, USA: Springer-Verlag New York, Inc., 1999, pp. 459474.
[14] C. K. Wong, M. Gouda, and S. S. Lam. Secure Group Communications
Using Key Graphs. ACM SIGCOMM, 1998.
[15] A. Perrig, D. Song, and J.D. Tygar. ELK, A new protocol for Efficient
Large-group Key distribution. IEEE Security and Priavcy Symposium,
May 2001.
[16] M. Waldvogel, G. Caronni, D. Sun, N. Weiler, , and B. Plattner. The
VersaKey Framework : Versatile Group Key Management, IEEE
Journal on Selected Areas in Communications (Special Issues on
Middleware), 17(8):16141631, August 1999.
[17] T.Hardjono, B.cain, and L.Monga. Intra-domain Group key
Management for Multicast Security. IETF internet Draft, September
2000
[18] A.Ballardie. Scalable Multicast Key Distribution. May 1996,
RFC1949.
[19] S. Rafaeli and D. Hutchison. Hydra: a decentralized group key
management. 11th IEEE International WETICE: Enterprise Security
Workshop, June 2002.
[20] Suvo Mittra, Iolus: A Framework for Scalable Secure Multicasting,
ACM SIGCOMM, 1997
[21] MOLVA, R. AND PANNETRAT, A. 1999. Scalable multicast security
in dynamic groups. In Proceedings of the 6th ACMConference on
Computer and Communications Security. (Singapore, Nov.). ACM, New
York, 101112.
[22] S.Setia, S.Koussih, S.Jaodia, and E.Harder. Kronos: A scalable Group
Re-Keying Approach for Secure Multicast. Proc. of IEEE Symposium
on Security and Privacy,2000
[23] Y.R. Yang, X.S. Li, X.B. Zhang, and S.S. Lam. Reliable Group
Rekeying: A Performance Analysis. TR-01-21, June 2001.
[24] KIM, Y., PERRIG, A., AND TSUDIK, G. 2000. Simple and fault-
tolerant key agreement for dynamic collaborative groups. In Proceedings
of the 7th ACM Conference in Computer and Communication Security,
(Athens, Greece Nov.). (S. Jajodia and P. Samarati, Eds.), pp. 235241.
[25] F.Staddon, S.Miner, M.Franklin, D.Balfanz, and D.Dean. Selfhealing
Key Distribution with Revocation. In Proc. Of the IEEE Symposium on
Security and Privacy, Oakland, CA, May 2002.
[26] Alireza Nemaney Pour, Kazuya Kumekawa, Toshihiko Kato, Shuichi
Itoh, A Hierarchical group key management scheme for secure
multicast increasing efficiency of key distribution in leave operation,
Elsevier,Computer Networks, August 2007.
AUTHORS PROFILE

1
R.Siva Ranjani is a research scholar in Andhra
University under the supervision of Prof.P.S.Avadhani and
Dr.D.Lalitha Bhaskari in Computer Science and Systems
Engineering. She received her M.Tech (CSE) from Andhra
University and presently working as Associate Professor in
CSE Department of GMRIT. She is a Life Member of
ISTE. Her research areas include Network Securiy,
Cryptography, Group Key Managment.
2
Mrs. Dr. D. Lalitha Bhaskari is an Associate Professor in
the department of Computer Science and Engineering of
Andhra University. She is guiding more than 8 Ph. D
Scholars from various institutes. Her areas of interest
include Theory of computation, Data Security, Image
Processing, Data communications, Pattern Recognition.
Apart from her regular academic activities she holds
prestigious responsibilities like Associate Member in the Institute of
Engineers, Member in IEEE, Associate Member in the Pentagram Research
Foundation, Hyderabad, India.
3
Dr. P. S. Avadhani is a Professor in the department of
computer Science and Engineering of Andhra University.
He has guided 7 Ph. D students, 3 students already
submitted and right now he is guiding 12 Ph. D Scholars
from various institutes. He has guided more than 100
M.Tech. Projects. He received many honors and he has
been the member for many expert committees, member
of Board of Studies for various universities, Resource person for various
organizations. He has co-authored 4 books. He is a Life Member in CSI,
AMTI, ISIAM, ISTE, YHAI and in the International Society on Education
Technology. He is also a Member of IEEE, and a Member in AICTE.

Vol. 2, No. 11, 2011
87 | P a g e
CluSandra: A Framework and Algorithm for Data
Stream Cluster Analysis

Jose R. Fernandez
Department of Computer Science
Pensacola, FL, USA
Eman M. El-Sheikh
Department of Computer Science
Pensacola, FL, USA

AbstractThe clustering or partitioning of a datasets records
into groups of similar records is an important aspect of
knowledge discovery from datasets. A considerable amount of
research has been applied to the identification of clusters in very
large multi-dimensional and static datasets. However, the
traditional clustering and/or pattern recognition algorithms that
have resulted from this research are inefficient for clustering
data streams. A data stream is a dynamic dataset that is
characterized by a sequence of data records that evolves over
time, has extremely fast arrival rates and is unbounded. Today,
the world abounds with processes that generate high-speed
evolving data streams. Examples include click streams, credit
card transactions and sensor networks. The data streams
inherent characteristics present an interesting set of time and
space related challenges for clustering algorithms. In particular,
processing time is severely constrained and clustering algorithms
must be performed in a single pass over the incoming data. This
paper presents both a clustering framework and algorithm that,
combined, address these challenges and allows end-users to
explore and gain knowledge from evolving data streams. Our
approach includes the integration of open source products that
are used to control the data stream and facilitate the harnessing
of knowledge from the data stream. Experimental results of
testing the framework with various data streams are also
discussed.
Keywords-data stream; data mining; cluster analysis; knowledge
discovery; machine learning; Cassandra database; BIRCH;
CluStream; distributed systems.
I. INTRODUCTION
According to the International Data Corporation (IDC), the
size of the 2006 digital universe was 0.18 zettabytes
1
and the
IDC has forecasted a tenfold growth by 2011 to 1.8 zettabytes
[17]. One of the main sources of this vast amount of data are
streams of high speed and evolving data. Clustering analysis is
a form of data mining whose application has, relatively
recently, started to be applied to data streams. The unbounded
and evolving nature of the data that is produced by the data
stream, coupled with its varying and high-speed arrival rate,
require that the data stream clustering algorithm embrace these
properties: efficiency, scalability, availability, and reliability.
One of the objectives of this work is to produce a distributed
framework that addresses these properties and, therefore,
facilitates the development of data stream clustering
algorithms for this extreme environment. Another objective is

1
One zettabyte equals 10
21
bytes or one billion terabytes.
to implement a clustering algorithm that is specifically
designed to leverage the distributed framework. This paper
describes that clustering algorithm and the distributed
framework, which is entirely composed of off-the-shelf open
source components. The framework is referred to simply as
CluSandra, while the algorithm, which is deployed onto the
framework, is referred to as the CluSandra algorithm.
CluSandras primary pillars are a database system called
Cassandra [9][15] and a message queuing system (MQS).
Cassandra, which is maintained by the Apache Software
Foundation (ASF), is a new breed of database system that is
referred to as a NoSQL database. At its core, Cassandra is a
distributed hash table (DHT) designed to tackle massive
datasets, perform in near-time and provide linear scalability
[9]. The MQS can be any number of either open source or
commercial message queuing systems that implement the Java
Message Service (JMS) API. All experimentation, related to
this work, was performed using the Apache ActiveMQ [16]
queuing system.
The combination of the CluSandra framework and
algorithm provides a distributed, scalable and highly available
clustering system that operates efficiently within the severe
temporal and spatial constraints associated with real-time
evolving data streams. Through the use of such a system, end-
users can also gain a deeper understanding of the data stream
and its evolving nature in both near-time and over different
time horizons.
A. Data Stream
A data stream is an ordered sequence of structured data
records with these inherent characteristics: fast arrival rate,
temporally ordered, evolves over time, and is unbounded [8].
The data streams arrival rate can be in the order of thousands
of data records per second, the concepts that are derived from
the data stream evolve at varying rates over time and, because
the data stream is unbounded, it is unfeasible to store all of its
records in any form of secondary storage (e.g., DBMS). The
data streams evolutionary characteristic is referred to as
concept drift [3]. This type of change may come as a result of
the changing environment of the problem; e.g., floating
probability distributions, migrating clusters of data, loss of old
and appearance of new classes and/or features, class label
swaps, etc. [20] Examples of data streams include IP network
traffic, sensor networks, wireless networks, radio frequency
identification (RFID), customer click streams, telephone
records, etc. Today, there are many applications whose data is
best modeled as a data stream and not as a persistent set of
Vol. 2, No. 11, 2011
88 | P a g e
tables. The following are some examples of applications for
data stream processing [11]:
- Real-time monitoring of information systems that
generate vast amounts of data. For example, computer
network management, telecommunications call
analysis, internet applications (e.g., Google, eBay,
recommendation systems, click stream analysis) and
monitoring of power plants.
- Generic software for applications based on streaming
data. For example, finance (fraud detection, stock
market analysis), sensor networks (e.g., environment,
road traffic, weather forecasting, electric power
consumption).
In this paper, a data stream S is treated as an unbounded
sequence of pairs (s,t) , where s is a structured data record (set
of attributes) and t is a system-generated timestamp attribute
that specifies when the data record was created. Therefore, t
may be viewed as the data streams primary key and its values
are monotonically increasing [7]. The timestamp values of one
data stream are independent from those of any other data
stream that is being processed within CluSandra. Since data
streams comprise structured records, streams comprising
unstructured data (e.g., audio and video streams) are not
considered data streams within the context of this paper.
B. Cluster Analysis
Cluster analysis or clustering is a process by which similar
objects are partitioned into groups. That is, all objects in a
particular group are similar to one another, while objects in
different groups are quite dissimilar. The clustering problem is
formally defined as follows: for a given set of data points, we
wish to partition them into one or more groups of similar
objects, where the notion of similarity is defined by a distance
function [21]. Clustering is a very broad topic that lies at the
intersection of many disciplines such as statistics, machine
learning, data mining, and linear algebra [12]. It is also used
for many applications such as pattern recognition, fraud
detection, market research, image processing, and network
analysis.
The focus of this work is on data clustering, which is a
type of data mining problem. Large multi-dimensional datasets
are typically not uniformly distributed. By identifying the
sparse and dense areas of the data space, data clustering
uncovers the distribution patterns of the dataset [10]. In
general, data clustering seeks to partition unlabeled data
records from a large dataset into labeled clusters, which is a
form of classification. Classification is an important problem
that has been studied extensively within the context of data
streams[3][4]. With respect to evolving data streams,
clustering presents an attractive advantage, because it is easily
adapted to changes in the data and can, therefore, be used to
identify features that distinguish different clusters [12]. This is
ideal for concept drifting data streams.
There are different data types (e.g., binary, numerical,
discrete) that need to be taken into account by data clustering
algorithms; the CluSandra algorithm only processes numerical
data. Future work can deploy additional algorithms, designed
to handle other data types, onto the CluSandra framework.
This work also assumes that the values for all the data records
attributes are standardized; therefore, there is no preprocessing
of the data records. Numerical data are continuous
measurements of a roughly linear scale [12]. When working
with data records whose attributes are of this data type, the
records can be treated as n-dimensional vectors, where the
similarity or dissimilarity between individual vectors is
quantified by a distance measure. There are a variety of
distance measures that can be applied to n-dimensional
vectors; however, the most common distance measure used for
continuous numerical data is the Euclidean measure:
d(i,j) = (1)
where x
ik
and x
jk
are the k
th
variables for the n-dimensional
data records i and j. For example, suppose you have two 2-
dimenisonal data records as follows: (1,3) and (4,1). The
Euclidean distance between these two records is the following:
= 3.60 (2)
If the data streams records are viewed as Euclidean
vectors in Euclidean n-space, the distance between any two
vectors (records) is the length of the line connecting the two
vectors tips or points. The lower the resulting value, the
closer (similar) the two vectors. The CluSandra algorithm
utilizes Euclidean distance as a measure to determine how
similar or close a new data record is to a clusters centroid
(mean). It is also used to find the distance between two
clusters centroids.
The next section discusses related work in this area.
Section III describes the CluSandra framework and Section IV
describes the cluster query language that was developed.
Section V presents the experimental results and section VI
discusses the conclusions and opportunities for future work.
II. RELATED WORK
A considerable amount of research has been applied to
clustering very large multi-dimensional datasets. One of the
key challenges, which has been the subject of much research,
is the design of data clustering algorithms that efficiently
operate within the time and space constraints presented by
very large datasets. That is, the amount of available memory
and required I/O time relative to the datasets size. These
constraints are greatly amplified by the data streams
extremely fast arrival time and unbounded nature. The
research work done in [5], [6], and [10] introduced concepts,
structures and algorithms that made great strides towards
efficient clustering of data streams. The CluSandra algorithm
is based on and expands on this work, as described below.
A. BIRCH
The CluSandra algorithm is based on the concepts and
structures introduced by the Balanced Iterative Reducing and
Clustering (BIRCH) [10] clustering algorithm. It is based on
the K-means (center-based) clustering paradigm and,
therefore, targets the spherical Gaussian cluster. K-means
provides a well-defined objective function, which intuitively
(x
ik
x
jk
)
2
k1
n
(1 4)
2
+(3 1)
2
Vol. 2, No. 11, 2011
89 | P a g e
coincides with the idea of clustering [19]. The simplest type of
cluster is the spherical Gaussian [19]. Clusters that manifest
non-spherical or arbitrary shapes, such as correlation and non-
linear correlation clusters, are not addressed by the BIRCH
and CluSandra algorithms. However, the CluSandra
framework does not preclude the deployment of algorithms
that address the non-spherical cluster types.
BIRCH mitigates the I/O costs associated with the
clustering of very large multi-dimensional and persistent
datasets. It is a batch algorithm that relies on multiple
sequential phases of operation and is, therefore, not well-
suited for data stream environments where time is severely
constrained. However, BIRCH introduces concepts and a
synopsis data structure that help address the severe time and
space constraints associated with the clustering of data
streams. BIRCH can typically find a good clustering with a
single pass of the dataset [10], which is an absolute
requirement when having to process data streams. It also
introduces two structures: cluster feature (CF) and cluster
feature tree. The CluSandra algorithm utilizes an extended
version of the CF, which is a type of synopsis structure. The
CF contains enough statistical summary information to allow
for the exploration and discovery of clusters within the data
stream. The information contained in the CF is used to derive
these three spatial measures: centroid, radius, and diameter.
All three are an integral part of the BIRCH and CluSandra
algorithms. Given N n-dimensional data records (vectors or
points) in a cluster where i = {1,2,3, , N}, the centroid ,
radius R, and diameter D of the cluster are defined as
(3)
(4)
(5)
where and are the i
th
and j
th
data records in the cluster
and N is the total number of data records in the cluster. The
centroid is the clusters mean, the radius is the average
distance from the objects within the cluster to their centroid,
and the diameter is the average pair-wise distance within the
cluster. The radius and diameter measure the tightness or
density of the objects around their clusters centroid. The
radius can also be referred to as the root mean squared
deviation or RMSD with respect to the centroid. Typically,
the first criterion for assigning a new object to a particular
cluster is the new objects Euclidean distance to a clusters
centroid. That is, the new object is assigned to its closest
cluster. Note, however, that equations 3, 4 and 5 require
multiple passes of the data set. Like BIRCH, the CluSandra
algorithm derives the radius and centroid while adhering to the
single pass constraint. To accomplish this, the algorithm
maintains statistical summary data in the CF. This data is in
the form of the total number of data records (N), linear sum,
and sum of the squares with respect to the elements of the n-
dimensional data records. This allows the algorithm to
calculate the radius, as follows:
(6)
For example, given these three data records: (0,1,1),
(0,5,1), and (0,9,1), the linear sum is (0,15,3), the sum of the
squares is (0,107,3) and N is 3. The CF, as described in
BIRCH, is a 3-tuple or triplet structure that contains the
aforementioned statistical summary data. The CF represents a
cluster and records summary information for that cluster. It is
formally defined as
(7)
where N is the total number of objects in the cluster (i.e., the
number of data records absorbed by the cluster), LS is the
linear sum of the cluster and SS is the clusters sum of the
squares:
(8)
(9)
It is interesting to note that the CF has both the additive
and subtractive properties. For example, if you have two
clusters with their respective CFs, CF
1
and CF
2
, the CF that is
formed by merging the two clusters is simply CF
1
+ CF
2
.
B. CluStream
CluStream [5] is a clustering algorithm that is specifically
designed for evolving data streams and is based on and extends
the BIRCH algorithm. This section provides a brief overview
of CluStream and describes the problems and/or constraints
that it addresses. This section also describes how the
CluSandra algorithm builds upon and, at the same time,
deviates from CluStream.
One of CluStreams goals is to address the temporal
aspects of the data streams single-pass constraint. For
example, the results of applying a single-pass clustering
algorithm, like BIRCH, to a data stream whose lifespan is 1 or
2 years would be dominated by outdated data. CluStream
allows end-users to explore the data stream over different time
horizons, which provides a better understanding of how the
data stream evolves over time. CluStream divides the
clustering process into two phases of operation that are meant
to operate simultaneously. The first, which is referred to as the
online phase, efficiently computes and stores data stream
summary statistics in a structure called the microcluster. A
microcluster is an extension of the BIRCH CF structure
x
0
x
0

x
i
i1
N
N
R
x
i
x
0
( )
2
i1
N
N
D
x
i
x
j
( )
2
j1
N
i1
N
N N 1
( )
x
i
x
j
x
i
2
x
i ( )
2
/ N
( )
1/ N 1
( )
CF N, LS, SS
LS x
i
i1
N
SS x
i
2
i1
N

Vol. 2, No. 11, 2011
90 | P a g e
whereby the CF is given two temporal dimensions. The second
phase, which is referred to as the offline phase, allows end-
users to perform macroclustering operations on a set of
microclusters. Macroclustering is the process by which end-
users can explore the microclusters over different time
horizons. To accomplish this, CluStream uses a tilted time
frame model for maintaining the microclusters. The tilted time
frame approach stores snapshots (sets) of microclusters at
different levels of granularity based on elapsed time. In other
words, as time passes, the microclusters are merged into
coarser snapshots. The CluSandra algorithm is based upon and
extends CluStream and the design of the CluSandra
framework is based on the concepts of both microclustering
and macroclustering. However, the general approach taken by
CluSandra for these two operational phases is quite different
than that taken by CluStream.
CluStreams microcluster extends the CF structure by
adding two temporal scalars or dimensions to the CF. The first
scalar is the sum of the timestamps of all the data records that
have been absorbed by the cluster and the second is the sum of
the squares of the timestamps. Thus the CF triplet, as defined
by BIRCH, is extended as follows:
(10)
Where ST is the sum of the timestamps and SST is the sum of
the squares of the timestamps. Note that this extended CF
retains its additive and subtractive properties. From
henceforth, this extended version of the CF is simply referred
to as a microcluster. The ST and SST can be applied to
expression (6) to arrive at the temporal standard deviation of
the microcluster as follows:
(11)
CluStreams microclustering process collects and
maintains the statistical information in such a manner that the
offline macroclustering phase can make effective use of the
information. For example, macroclustering over different time
horizons and exploring the evolution of the data stream over
these horizons.
CluStream defines a fixed set of microclusters that it
creates and maintains in-memory. This set, M = {M
1
, M
2
, ,
M
q
} is the current online working set with q being the
maximum number of microclusters in the set and each
microcluster in M being given a unique id. The CluStream
paper states that the algorithm is very sensitive to any further
additions to the microcluster snapshot, as this negatively
impacts q. This type of space constraint is one that CluSandra
removes by relying on Cassandra to serve as a highly reliable
and scalable real-time distributed cluster database. Like
CluStream, the CluSandra algorithm also maintains a working
set of microclusters in memory; however, the size of this
working set is dictated by the size of a sliding time window in
combination with a maximum boundary threshold (MBT).
CluStreams updating of the microclusters is similar to that
of BIRCH. When a new data record is presented by the data
stream, it is assigned to the closest microcluster Mi in M. After
finding the closest microcluster Mi, it is then determined if Mi
can absorb the new record without exceeding its MBT.
CluStream defines the MBT as, a factor of t of the RMSD of
the data points in M
i
from the centroid. This is another way
of referring to the standard deviation of the cluster times a
factor t to arrive at the final maximum radius. The
microcluster M
i
is updated whenever it absorbs a new data
record. If M
i
has absorbed only one data record, the RMSD
cannot be calculated; therefore, for those clusters having
absorbed only one data record, the MBT is derived as the
distance from M
i
to its closest neighbor times a factor r. The
CluSandra algorithm uses a similar approach in locating the
closest microcluster. However, it does not use the nearest
neighbor approach for those instances where the closest
microcluster has absorbed only one entry. It instead
determines if the closest microcluster can absorb the data
record without exceeding a configurable and fixed maximum
radius, which is CluSandras MBT.
If M
i
cannot absorb the new data record, CluStream creates
a new microcluster to host the data record. To conserve
memory, CluStream requires that either one of the existing
microclusters in M be deleted or two microclusters be merged.
Only those microclusters that are determined to be outliers are
removed from M. If an outlier cannot be found in M, two
microclusters are merged. The CluSandra algorithm deviates
from this approach since it simply adds a new microcluster to
its current working set. Again, the size of CluSandras in-
memory working set is managed according to a temporal
sliding window and the specified MBT. Any microclusters
that are no longer active within the current sliding window are
removed from the working set, but not before being persisted
to the Cassandra cluster database.
The temporal scalars (ST, SST) of the microcluster, in
combination with a user-specified threshold o, are used to look
for an outlier microcluster in M. The ST and SST scalars
allows CluStream to calculate the mean and standard deviation
of the arrival times of the data records in Ms microclusters.
CluStream assumes that the arrival times adhere to a normal
distribution. With the mean and standard deviation of the
arrival times calculated, CluStream calculates a relevance
stamp for each of the microclusters. A microcluster whose
relevance stamp is less than the threshold o is considered an
outlier and subject to removal from M. If all the relevance
stamps are recent enough, then it is most likely that there will
be no microclusters in M whose relevance stamp is less than o.
If and when this occurs, CluStream merges the two closest
microclusters in M and assigns the resulting merged
microcluster a listid that is used to identify the clusters that
were merged to create this new merged microcluster. So as
time progresses, one microcluster may end up comprising
many individual microclusters.
Unlike BIRCH, CluStream does not utilize a tree structure
to maintain its microclusters. At certain time intervals, and
while M is being maintained as described above, M is
persisted to secondary storage. Each instance of M that is
persisted is referred to as a snapshot. CluStream employs a
logarithmic based time interval scheme, which is referred to as
CF N, LS, SS, ST, SST
SST ST
( )
2
( )
1/ N 1
( )
Vol. 2, No. 11, 2011
91 | P a g e
a pyramidal time frame, to store the snapshots. This technique
guarantees that all individual microclusters in M are persisted
prior to removal from M or being merged with another
microcluster in M. This allows a persisted and merged
microcluster (i.e., those having a listid) in a snapshot to be
broken down (via the microclusters subtractive property) into
its constituent (individual), finer-grained microclusters during
the macroclustering portion of the process. The opposite is
also available, whereby the additive property allows finer-
grained/individual and merged microclusters to be merged into
more course grained microclusters that cover specified time
horizons. Snapshots are classified into orders, which can vary
from 1 to log
2
(T), where T is the amount of clock time elapsed
since the beginning of the stream[5]. The number of snapshots
stored over a period of T time units is
(12)
For example, if o = 2 and the time unit or granularity is 1
second, then the number of snapshots maintained over 100
years is as follows:
(13)
The CluSandra algorithm does not operate within the same
memory constraints as CluStream. During the CluSandra
algorithms microclustering process, microclusters are not
merged to accommodate a new microcluster and there is no
need to search for possible outliers that can be targeted for
removal from M. When a new microcluster is created, it is
simply added to the current in-memory working set and
persisted to the Cassandra cluster database. Also, when a
microcluster absorbs a data record, the microcluster is
immediately persisted to the cluster database; there is no
dependence on a periodic time interval scheme for persisting
M to secondary storage.
CluStream does not include a database system of any kind
for persisting its snapshots and does not fully address the
transactional integrity associated with snapshot persistence.
After the in-memory snapshot M has been updated, there may
exist a relatively lengthy time period before the snapshot is
persisted to the local file system. This raises the risk that the
snapshots state may be lost due to process, system or
hardware failure. Also, CluStream does not address Ms
recoverability. If the machine fails and is restarted, CluStream
cannot reliably recover Ms state prior to the failure. In other
words, there is no transactional integrity associated with Ms
persistence. One of the goals of the CluSandra framework is to
introduce the necessary components that guarantee the
transactional integrity of microcluster persistence.
III. CLUSANDRA FRAMEWORK
Figure 1 is a level 1 context-level data flow diagram[2]
(DFD) that represents, at a high-level, the CluSandra
framework. The frameworks core components are written
entirely in version 1.5 of the Java programming language;
however, the framework supports client components that are
written in a variety of programming languages. The CluSandra
framework and algorithm are based on temporal and spatial
aspects of clustering. The algorithm, which is implemented in
a MicroClustering Agent (MCA) framework component and
described in more detail in section D, uses temporal and
spatial measures (radius, distance) to group the data streams
records into microclusters. The microclusters are stored in the
Cassandra database and later accessed by offline processes
(e.g., aggregator) and/or end-users. A Cluster Query Language
(CQL) is provided by the framework to facilitate the querying
and analysis of the microclusters in the database. As
previously noted, Cassandra is an implementation of a DHT
that comprises two or more distributed machines configured in
a peer-to-peer ring network topology. The ring of machines or
nodes is called a Cassandra cluster. To meet the most
demanding environments, Cassandra can elastically scale up
from a one or two node cluster to a cluster comprising 10s, if
not 100s, of nodes and then back down to one or two nodes.
Each node in the cluster may also optionally host the
CluSandra frameworks other executable components.
A. StreamReader
The StreamReader component is responsible for reading
the data streams structured data records, wrapping those data
records in a CluSandra-specific DataRecord object, and then
sending the DataRecords to the CluSandra message queuing
system (MQS), where they are temporarily stored or buffered
for subsequent processing. The CluSandra framework
automatically time stamps the DataRecords when they are
created, but the StreamReader can also override the
frameworks timestamp. This may be required if, for example,
the raw data stream records are already time-stamped. The
time stamps are critically important, because they are used to
record the data streams timeline. The raw data stream record
that is read by the StreamReader is treated as a
multidimensional vector containing one or more continuous
numerical values. This vector is encapsulated by the
DataRecord object. It is critical for the StreamReader to keep
up with the data streams arrival rate, which is assumed to be
in the 1000s of records per second
2
. However, this should not
be an issue given the simple and straightforward nature of the
StreamReaders main purpose.

2
It is also quite possible that the StreamReader is the stream
generator.
+1
( )
log
2
T
( )
2+1
( )
log
2
100 365 24 60 60
( )
95
Vol. 2, No. 11, 2011
92 | P a g e

Figure 1. CluSandra data flow diagram
CluSandras MQS provider is remotely accessed via a
TCP/IP network; therefore, the StreamReader component
does not have to reside on the same node as the MQS provider
nor does it have to reside in a Cassandra cluster node. The
StreamReader can even be embedded within the component or
device (e.g., sensor or router) that produces the data stream.
StreamReaders can be distributed across a population of such
devices, all writing to the same or different MQS queue, with
each queue dedicated to a particular data stream. Most MQS
providers, like ActiveMQ, support clients written in a variety
of languages. If a StreamReader is not written in the Java
programming language and would like to take advantage of
the component that implements the CluSandra algorithm, then
it requires a transformational component; in other words, a
Java proxy that can create a Java DataRecord object and place
it in the MQS for the non-Java StreamReader.
B. Timeline Index
The timeline index (TI), which is not depicted in the DFD,
represents the data streams timeline. It is an important
temporal-based index that is maintained in the Cassandra
database and used for maintaining the clusters in the database.
In the Cassandra vernacular, the TI is a type of secondary
index. There is one TI defined for each data stream that is
being processed within the CluSandra framework. The TI is
implemented as a Cassandra SuperColumnFamily. At an
abstract level, each entry in the TI represents a second in the
data streams lifespan. Each entrys contents or value is a
Cassandra SuperColumn whose entries (Columns) form a
collection or set of row keys to clusters that were or are still
active during that second in time. A data streams timeline can
run for any amount of time. During development and testing,
the streams timeline may extend over a handful of seconds,
while in production environments, the timeline may extend
over months, if not years.
The TIs implementation comprises a Cluster Index Table
(CIT). The CIT includes a row for each day of the year and
each row comprises 86400 SuperColumns; one SuperColumn
for each second of that particular day. Each SuperColumn
contains one or more Columns whose values are keys to
clusters in the Cluster Table (CT). Also, a SuperColumn can
be assigned different names from row-to-row, and in this case
the name of a SuperColumn is a timestamp that corresponds
to a second for that day. Each row key of the CIT is given a
value that corresponds to the zeroth second of a particular
day. The motivation behind the TIs implementation is that
Cassandra is not currently set up to perform well with sorted
rows and returning ranges within those sorted rows. However,
it is set up to sort columns and return ranges or slices of
columns, based on their names.
C. Message Queuing System
The message queuing system (MQS) is a critical piece of
the CluSandra framework. It serves as a reliable asynchronous
message store (buffer) that guarantees delivery of its queued
messages (DataRecords). So as a dam is used to control a wild
raging river and harness it to produce electricity, so is the
MQS used to control the evolving high-speed data stream so
that it can be harnessed to produce knowledge. In CluSandras
case, a MQS queue serves as the streams dam and its
contents of time-stamped DataRecords form the reservoir. The
CluSandra framework supports the simultaneous processing
of many data streams; therefore, there may very well be many
queues defined within the MQS; one queue for each data
stream.
StreamReader Data Stream
Message Queuing
System
Data Records
Time Stamped
Data Records
Time Stamped
Data Records
Cassandra
Data Store
MicroClusters
MicroClustering
Agents
SuperClusters
MicroClusters
MacroClustering
Queries
Vol. 2, No. 11, 2011
93 | P a g e
The primary motivation for having the CluSandra
framework incorporate a MQS is to control the data stream.
More precisely, the MQS provides a reliable DataRecord store
that temporarily buffers and automatically distributes
DataRecords across one or more instances of the
microclustering agent (MCA) component. A group of
identical MCAs that consume DataRecords from the same
queue is called a swarm. Many swarms can be defined and
distributed across the framework, with each swarm reading
from its unique queue. The queue is capable of retaining
thousands of DataRecords and guarantees their delivery to the
swarm, which is responsible for ensuring that the number of
DataRecords in the MQS queue is maintained at an acceptable
level. The swarm size is, therefore, a function of the data
stream rate. Data streams with extremely fast arrival rates
produce very large volumes of DataRecords and require a
correspondingly large swarm.
The CluSandra framework is designed to utilize any MQS
that implements the Java Message Service (JMS) API. The
vast majority of MQS providers, both open source and
commercial, implement the JMS. The JMS is an industry
standard Java messaging interface that decouples applications,
like the StreamReader, from the different JMS
implementations. JMS thus facilitates the seamless porting of
JMS-based applications from one JMS-based queuing system
to another. Within the context of the JMS, the process that
places messages in the MQSs queues is called the producer,
the process that reads messages from the queue is called the
consumer, and both are generically referred to as clients. So
the StreamReader is a producer, the MCA is a consumer and
they are both clients.
The MQS guarantees delivery of DataRecords to the
swarm. This guarantee means that DataRecords are delivered
even if the MQS or machine hosting the MQS were to fail and
be restarted. The MQS achieves this guarantee via a
combination of message persistence and a broker-to-client
acknowledgment protocol. These MQSs can be configured to
persist their messages to either file systems (distributed or
local) or database management systems (DBMS). For the
CluSandra framework, the MQS is configured to use a
distributed or shared file system and not a DBMS. The file
system provides much better throughput performance than
does a DBMS.
These MQS systems are also architected to provide fault-
tolerance via redundancy. For example, you can run multiple
message broker processes across multiple machines, where
certain message brokers can act as hot or passive standbys for
failover purposes. The message broker is the core component
of the MQS and is the component responsible for message
delivery. One overarching requirement is to ensure that this
level of reliability and/or fault tolerance be an inherent quality
of the CluSandra framework. These MQS systems include
many features, but it is beyond the scope of this paper to list
all the features.
D. Microclustering Agent
The microclustering agent (MCA) is the framework
component that consumes DataRecords from a particular
MQS queue and implements a clustering algorithm that
produces microclusters. This section describes the MCA that
implements the CluSandra algorithm and is delivered with the
CluSandra framework.
Like CluStream, the CluSandra algorithm tackles the one-
pass data stream constraint by dividing the data stream
clustering process into two operating phases: online and
offline. Microclustering takes place in real-time, computing
and storing summary statistics about the data stream in
microclusters. There is also an optional offline aggregation
phase of microclustering that merges temporally and spatially
similar microclusters. Macroclustering is another offline
process by which end-users create and submit queries against
the stored microclusters to discover, explore and learn from
the evolving data stream. As CluStream extended the BIRCH
CF data structure, so does the CluSandra algorithm extend the
CluStreams CF structure as follows:
(14)
The CFT is used to represent either a microcluster or
supercluster in the CluSandra data store. The term cluster
applies to both micro and superclusters. The CT and LAT
parameters are two timestamp scalars that specify the creation
time and last absorption time of the cluster, respectively.
More precisely, CT records the time the cluster was created
and the LAT records the timestamp associated with the last
DataRecord that the cluster absorbed. When a microcluster is
first created, to absorb a DataRecord that no other existing
microcluster can absorb, both the CT and LAT parameters are
assigned the value of the DataRecords timestamp. All
timestamps in CluSandra are measured in the number of
milliseconds that have elapsed since Unix or Posix epoch
time, which is January 1,1970 00:00:00 GMT. The IDLIST is
also new and is used by superclusters.
Over time, a particular pattern in the data stream may
appear, disappear and then reappear. This is reflected or
captured by two microclusters with identical or very similar
spatial values, but different temporal values. The time horizon
over which a cluster was active can be calculated by the CT
and LAT parameters and more detailed statistical analysis can
be performed based on the other temporal, as well as spatial
parameters. For example, the temporal density of the
DataRecords and their spatial density with respect to one
another and/or their centroid. An inactive microcluster is no
longer capable of absorbing data records; however, during
macroclustering, it can be merged with other inactive and
active microclusters to form a supercluster. The algorithms
microclustering phase, therefore, works within a specified
temporal sliding window and updates only those microclusters
that are within that time window. Such a temporal sliding
window is required, because of the unbounded nature of the
data stream.
As previously mentioned, the MCA consumes sets of
DataRecords from its assigned MQS queue. The framework
provides a read template for the MCA that includes this
functionality; therefore, the one implementing the MCAs
clustering algorithm does not need to concern herself with this
functionality. The members of a set D of DataRecords are
consumed, by the read template, in the same order that they
CFT N, LS, SS, ST, SST, CT, LAT, IDLIST
Vol. 2, No. 11, 2011
94 | P a g e
were produced by the StreamReader and the temporal order of
the DataRecords is maintained by the MQS. The set D can,
therefore, be viewed as a temporal window of the data stream.
D = {d
1
, d
2
, , d
k
} (15)
The maximum number of DataRecords in D is
configurable. Also, the amount of time the read template
blocks on a queue, waiting to read the maximum number of
DataRecords in D, is configurable. The read template
processes the set D whenever the maximum number has been
reached or the read time expires and D = C. The read time
should be kept at a relatively high value. So, 0 < k <= m,
where k is the total number of DataRecords in D and m is the
maximum. If the read time expires and D = C, the MCA
simply goes back to blocking on the queue for the specified
read time.
After the read template has read a set D from the queue, it
gives it to the clustering algorithm via a specified interface.
The following describes the CluSandra clustering algorithm
and from henceforth it is simply referred to as the CA. When
the CA receives a set D, it begins the process of partitioning
the DataRecords in D into a set M of currently active
microclusters. On startup, M is an empty set, but as time
progresses, M is populated with microclusters. The CA then
determines the time horizon h associated with D. The range of
h is R
h
and it is defined by the newest and oldest DataRecords
in D. Therefore, R
h
= {t
o
-t
e
,t
y
}, where t
e
is the configurable
microcluster expiry time and t
o
and t
y
represent the oldest and
newest DataRecords in D, respectively. All microclusters in
M, whose LAT does not fall within R
h
, are considered inactive
microclusters and removed from M. Thus the CA always
works within a sliding temporal window that is defined by R
h
.
When the CA completes the processing of D, it writes all new
and updated microclusters in M to the Cassandra data store,
gives control back to read template, and starts the partitioning
process over again when it is given a new set D of
DataRecords.
To partition the DataRecords in D, the CA iterates through
each DataRecord in D and selects a subset S of microclusters
from M, where all microclusters in S are active based on the
current DataRecords (d
c
) timestamp. For example, if the
timestamp for d
c
is t
d
, then only those microclusters in M
whose LAT value falls within the range, {t
d
-t
e
,t
d
} are added to
S. The value t
e
is the configurable microcluster expiry time.
Depending on the rate of the data stream and the microcluster
expiry time, it is possible that S = M and so, S _ M . The CA
uses the Euclidean distance measure to find the microcluster
in S that is closest to d
c
. The MBT is then used to determine if
the closest microcluster M
i
in S can absorb d
c
. The MBT is a
configurable numeric value that specifies the maximum radius
of a microcluster. Again, the radius or RMSD is the root mean
squared deviation of the cluster and is derived according to
(6). If M
i
can absorb d
c
without breaching the MBT, M
i
is
allowed to absorb d
c
and is placed back in M, else a new
microcluster is created to absorb d
c
and that new microcluster
is then added to M. If S = C, the CA simply creates a new
microcluster to absorb d
c
and adds the new microcluster to M.
One method for deriving a MBT value for the target data
stream is by sampling the data stream to derive an average
distance (measure of density) between DataRecords and then
using some fraction of that average density. If the CA has not
been assigned a MBT, it will derive the MBT based on the
first set of data records that it receives from the read template.
If M
i
has previously absorbed only one DataRecord (i.e.,
N=1), the RMSD cannot be calculated. In such a case, the
MBT is used to determine if the microcluster can absorb the
DataRecord.
Depending on the distribution of the data stream, as it
evolves over time, microclusters of all sizes appear, disappear,
and may reappear. The number and sizes of the microclusters
are a factor of not only the data streams evolutionary pattern,
but also of the MBT and microcluster expiry-time. The
smaller the MBT, the more microclusters will be produced
and vice versa. However, the optional offline aggregation
and/or macroclustering phases can be used to merge those
microclusters that are deemed similar. The next section
describes the CluSandra Aggregator component, which is
responsible for the aggregation and merges those
microclusters whose radii overlap. Please note that the
Aggregator does not produce superclusters; it simply merges
overlapping microclusters.
E. Aggregating Microclusters
If there is an MCA swarm distributed across the
CluSandra frameworks nodes, then it is very likely for the
swarm to create microclusters that are very similar, if not
equal, both temporally and spatially (see figure 2). This
occurs if two or more MCAs in the swarm process a set of
DataRecords with equal or overlapping time horizons (h
e
).
This may also occur as a natural side effect of the clustering
algorithm.

Figure 2. Overlapping clusters
If any microclusters temporally and spatially overlap, then
those microclusters may be viewed as one microcluster. Given
a two-dimensional vector-space, the figure above illustrates an
example where three MCAs have created three microclusters
(dots represent the microclusters centroids) that are so close
to one another, both spatially and temporally, that they should
be merged into one microcluster. The merging of these
microclusters is performed by an offline aggregator
component that sweeps through the data stream timeline (CIT)
performing such merges. Microcluster aggregation is a type of
agglomerative clustering procedure whereby individuals or
. .
.
Vol. 2, No. 11, 2011
95 | P a g e
groups of individuals are merged based upon their temporal
and spatial proximity to one another [13]. Agglomerative
procedures are probably the most widely used of the
hierarchical clustering methods [13]. The result of a merger,
performed by the aggregator, is immutable (see figure 3). This
aggregator, which is provided as part of the CluSandra
algorithm package, should not be confused with
superclustering and macroclustering, which is discussed in the
following section.

Figure 3. Before and after aggregation
The microcluster expiry time is used to determine if two
microclusters temporally overlap. To determine if two
temporally overlapping microclusters also spatially overlap,
the aggregator compares the distance between their centroids
with the sum of their radii. If the distance is less than the sum
of the radii, then the two microclusters spatially overlap.
There may even be instances where one microcluster is
entirely within the other. The aggregator is given a
configurable property called the overlapFactor. This property
is used to specify the amount of overlap that is required to
deem two microclusters similar enough to merge. The radii of
the microcluster that results from the merge may be greater
than the radii of the two merged microclusters; therefore, it
will occupy more space, as well as time and be capable of
absorbing more surrounding microclusters that temporally
overlap. So the less the overlapFactor, the greater the
probability of creating very large clusters that may mask out
interesting patterns in the data stream.
To minimize the occurrence of overlapping microclusters,
all members of a MCA swarm can work from the same or
shared set of microclusters in the data store. However, this
would have required coordination between the distributed
members of the swarm. Unfortunately, this level of
coordination requires a distributed locking mechanism that
introduces severe contention between the members of the
swarm. It is for this reason, that this approach of sharing
microclusters was not followed. Also, overlapping
microclusters is a natural side effect of the CluSandra
clustering algorithm.
F. Superclusters and Macroclustering
The CluSandra framework introduces the supercluster,
which is created when two or more clusters, either micro or
super, are merged via their additive properties. A supercluster
can also be reduced or even eliminated via its subtractive
property. The IDLIST in the CluSandra CFT is a collection or
vector of microcluster ids (Cassandra row keys) that identify a
superclusters constituent parts (i.e., microclusters). If the
IDLIST is empty, then it identifies the cluster as a
microcluster, else a supercluster.
Superclusters are created by the end-user during the
macroclustering process. Superclusters are created based on a
specified distance measure (similarity) and time horizon. Like
aggregation, superclustering is another type of agglomerative
clustering procedure; however, unlike traditional agglomertive
procedures where the result of a merger is immutable,
CluSandras superclusters can be undone. What makes this
possible is the subtractive property of the CFT. When a
supercluster is created, the earliest CT and LAT of its
constituent parts are used as the superclusters CT and LAT.
The CT and LAT thus identify the superclusters lifespan.
G. The Spring Framework
The open source Spring Framework [1] is relied upon by
the CluSandra framework for configuration and to reuse
Springs components. Except for Cassandra, all CluSandra
framework components are comprised of one or more Spring
POJOs (Plain Old Java Objects) that are configured via Spring
XML configuration files. The Spring Framework, which is
henceforth referred to simply as Spring, was created for the
specific purpose of minimizing the development costs
associated with Java application development. Spring
provides a number of Java packages whose components are
meant to be reused and that, in general, hide the complexities
of Java application development. However, at its core, Spring
is a modular dependency injection and aspect-oriented
container and framework. By using Spring, Java developers
benefit in terms of simplicity, testability and loose coupling.
The StreamReader and MCA components make heavy use
of Springs support for the JMS to send and receive
DataRecords to and from the MQS, respectively. Apache
ActiveMQ [16] has been selected as the CluSandra
frameworks default JMS provider (i.e., MQS). There are
many open source JMS providers or implementations.
ActiveMQ was chosen, because of its rich functionality,
beyond that specified by the JMS specification, and its robust
support for and integration with Spring.
As a JMS producer, the StreamReader indirectly uses
Springs JmsTemplate class to produce DataRecords destined
for the ActiveMQ message broker. The JmsTemplate is based
on the template design pattern and it is a convenience class
that hides much of the complexity of sending messages to a
JMS message broker. The JmsTemplate is integrated into a
send template (convenience class) provided by CluSandra;
therefore, the person implementing the StreamReader does not
have to concern herself with the implementation of this
functionality. This send template is similar to the previously
described read template for the MCA. In general, the use of
the JmsTemplate by both the send and read templates, in
combination with the corresponding Spring XML file, helps
ensure a decoupling between the CluSandra JMS clients and
whatever JMS provider is being used as the CluSandra MQS.
IV. CLUSTER QUERY LANGUAGE
The CluSandra framework includes a query language that
is used for querying the CluSandra algorithms cluster
. .
.
.
(a) Before aggrega on (b) A er aggrega on
Vol. 2, No. 11, 2011
96 | P a g e
database. The query language, which is referred to as the
cluster query language or CQL for short, includes the
following statements:
- Connect: This statement is used to connect to a particular
node in the Cassandra cluster.
- Use: This statement, which must be invoked immediately
after the Connect, is used to specify the keyspace to use
within the Cassandra cluster. A Cassandra keyspace is
analogous to a schema that is found in a relational database.
- Select: This statement, which is the one most often invoked,
is used for projecting clusters from the cluster store.
- Aggregate: This statement is used for invoking the
aggregator on all or a portion of the cluster database.
- Merge: This statement is used, as part of the offline
macroclustering process, to form superclusters.
- Sum: This is a relatively simple statement that is used to
return the total number of DataRecords that have been
absorbed by all the microclusters in the cluster database.
- Distance: This statement is used for acquiring the distance
between pairs of clusters.
- Overlap: This statement is used for acquiring the amount of
overlap between pairs of clusters or in other words, the
overlap percentage of the clusters radii.
CluSandras CQL is not to be confused with the Cassandra
Query Language, which is also referred to as CQL.
CluSandras CQL operates in either batch or interactive mode.
When invoked in batch mode, the user specifies a file that
contains CQL statements.
V. EMPIRICAL RESULTS
Several experiments were conducted to evaluate the
accuracy of the CluSandra algorithm, as well as the scalability
and reliability provided by the CluSandra framework. The
experiments are considered small-scale where the framework
comprised a cluster of 2-3 compute nodes. More large-scale
testing that comprises larger clusters of compute nodes is
planned for future work.
A. Test Environment and Datasets
Experiments designed to test the accuracy of the
CluSandra algorithm were conducted on an Intel Core i5 with
8 GB of memory running OS/X version 10.6.8. The
experiments included a real and synthetic dataset.
The real dataset was acquired from the 1998 DARPA
Intrusion Detection Evaluation Program, which was prepared
and managed by MIT Lincoln Labs. This is the same dataset
used for The Third International Knowledge Discovery and
Data Mining Tools Competition (KDD-99 Cup). According to
[5], this dataset was also used to run experiments on the
CluStream algorithm. The dataset is contained in a comma-
separated values (CSV) file where each line comprises a TCP
connection record; there are 4,898,431 records in the file. The
dataset records two weeks worth of normal network traffic,
along with bursts of different types of intrusion attacks, that
was simulated for a fictitious US military base. The attacks
fall into four main categories: DOS (denial-of-service), R2L
(unauthorized access from a remote machine), U2R
(unauthorized access to local super user privileges), and
PROBING (surveillance and other probing). Each connection
record contains 42 attributes, which provide information
regarding the individual TCP connection between hosts inside
and outside the fictitious military base. Some example
attributes are protocol (e.g., Telnet, Finger, HTTP, FTP,
SMTP, etc), duration of the connection, the number of root
accesses, number of bytes transmitted to and from source and
destination. The connection records are not time stamped. Of
the 43 attributes, 34 are of type continuous numerical. Every
record in the dataset is labeled as either a normal connection
or a connection associated with a particular attack. The
following is a list of all possible labels, with the number in
parenthesis being the total number of records in the dataset
having that particular label: back(2203), buffer_overflow(30),
ftp_write(8), guess_passwd(53), imap(1069), ipsweep(12481),
land(21), loadmodule(9), multihop(7), neptune(1072017),
nmap(2316), normal(972781), perl(3), phf(4), pod(264),
portsweep(10413), rootkit(10), satan(15892),
smurf(2807886), spy(2), teardrop(979), warezclient(1020),
warezmaster(20). Together, the normal, smurf and neptune
records comprise 99% of all records. The StreamReader
(KddStreamReader) for this experiment reads each connection
record and creates a DataRecord that encapsulates the 34
numerical attributes for that record. The KddStreamReader
then sends the DataRecord the frameworks MQS for
processing by the MCA, which in this case is an
implementation of the CluSandra algorithm. The
KddStreamReader includes a filter that is configured to read
all or any combination of records based on the label type. For
example, the end-user can configure the KddStreamReader to
process only the neptune records, only the smurf and neptune
records, or all the records.
The first series of experiments focused on processing the
more ubiquitous record types in the dataset. The first
experiment in the series had the KddStreamReader process
only the smurf records and assigned the MCA a time window
that captured the entire stream and an MBT (i.e., maximum
radius) of 1000; this same time window and MBT were
maintained throughout this series of experiments. The result
was one microcluster that had absorbed all 2,807,886 smurf
records and had a relatively dense radius of 242.34. This
experiment was executed five times with identical results. The
second experiment was identical to the first, except that the
KddStreamReader processed only the neptune records, which
is the second most ubiquitous record type. The result was,
once again, one microcluster that had absorbed all 1,072,017
neptune records, but with an even smaller radius of 103.15.
The next experiment targeted the normal record, which is the
third most ubiquitous record. One might expect that the result
would, once again, be only one microcluster. However, the
result was a set of 1,048 microclusters with very little to no
overlap between the microclusters and a high degree of
variance with respect to their radii and number of absorbed
DataRecords. This relatively large set of microclusters is to be
expected, because there exists a high degree of variance
within a set of normal connections. In other words, in a TCP
network, the usage across a set of normal connections is
typically not the same; the connections are being used for a
variety of different reasons. For example, some connections
are being used for email (SMTP), file transfer (FTP), and
Vol. 2, No. 11, 2011
97 | P a g e
terminal interfaces (Telnet, HTTP). It was also noted that it
took appreciably longer for this experiment to complete. This
is due to the specified time window encompassing the data
streams entire lifespan, which results in a much larger
number of microclusters in the in-memory working set. The
StreamReaders next target was the ipsweep records. The
result was one microcluster absorbing all but one ipsweep
record; the radius of this one microcluster was 145.66. The
distance between this microcluster and the one that had
absorbed the remaining ipsweep record was 53,340. It was,
therefore, assumed that this one neptune record was an outlier.
Finally, the KddStreamReader was focused on the portsweep
records. Unlike the previous attack-related experiments,
where only one or two microclusters were produced, this one
resulted in 24 microclusters; however, one microcluster
absorbed 90% of all the portsweep records and it had a radius
of 569. The other 23 microclusters had radii ranging in the
900-1000 range with very little overlap. Overall, the results of
this first series of experiments indicated a high level of
accuracy for the CluSandra algorithm. Also, in this series of
experiments, the average throughput rate for the
KddStreamReader was approximately 28,000 records per
second.
In the next series of experiments, the KddStreamReader
processed combinations of two or more connection record
types. To start, the KddStreamReader processed only the
neptune and smurf records, along with the same time window
and MBT as the previous series of experiments. The result
was two microclusters; the first having absorbed 2,807,639
records with a radius of 242.17, while the second absorbed
1,072,264 records with a radius of 101.52. It was clear that the
first and second microcluster accurately grouped the smurf
and neptune records, respectively. There was also no overlap
between the two microclusters, but it was interesting to see
that a relatively small number of smurf records were grouped
into the neptune microcluster. The KddStreamReader was
then configured to process all smurf, neptune and normal
records. The result was a set of 1,053 microclusters. The most
populated microcluster had absorbed 2,804,465 DataRecords
and had a radius of 232.62. This is clearly the smurf
microcluster, but note how it had lost approximately 3000
DataRecords. The second most populated microcluster had
absorbed 1,438,988 DataRecords and had a radius of 497.34.
This is clearly the neptune microcluster, but it had gained over
300,000 DataRecords and its radius had, as would be
expected, appreciably increased from 101.52. It was
concluded that a considerable number of normal DataRecords
had been absorbed by the neptune microcluster. This can be
addressed by reducing the MBT; however, this will also result
in an appreciably larger number of denser normal
microclusters. The third most populated microcluster had
absorbed 140,885 DataRecords with a radius of 544.89.
Population-wise, this is approximately one-tenth the size of
the neptune microcluster.
To test the accuracy of the sliding time window, the
windows size was reduced to capture the different bursts of
attacks. It was noted from the analysis of the dataset file that
the neptune and smurf attacks occur across a handful of
different bursts. So the size of the sliding time window was
reduced to 3 seconds, the MBT was kept at 1000, and the
StreamReader processed only the neptune records. The result
was 4 very dense microclusters (attacks) that had a radius of
100 or less, with 18 being the smallest radius. The first
microcluster absorbed 15 DataRecords and its lifespan
(duration) was only one second. The second microcluster was
created 4 seconds later, absorbed approximately 411,000
DataRecords and its lifespan was 12 seconds. Thus there was
a gap of 4 seconds, between the first and second microcluster,
where there were no neptune attacks. The third microcluster
was created 15 seconds after the second expired, absorbed
approximately 450,000 DataRecords, and its lifespan was 9
seconds. Finally, the fourth microcluster was created 4
seconds after the third expired, absorbed approximately
211,000 DataRecords and its lifespan was 6 seconds. The test
was repeated for the smurf connection records and the result
was one microcluster that had absorbed all the DataRecords
and had a lifespan of over one minute. The experiment was
run again, but with a window of 2 seconds. This time, the
result was 5 very dense microclusters; three had a lifespan of
1 or 2 seconds, one a lifespan of 12 seconds and the last a
lifespan of 42 seconds. These experiments proved the
accuracy of the sliding window.
The synthetic dataset was generated by a stream generator
that is loosely based on the Radial Basis Function (RBF)
stream generator that is found in the University of Waikatos
Massive Online Analysis (MOA) open source Java package
[8]. This RBF type of generator was used, because it produces
data streams whose data distribution adhere to a spherical
Gaussian distribution, which is the distribution that the
CluSandra algorithm is designed to process. During its
initialization, the RBF generator creates a set of randomly
generated centroids. The number of centroids in the set is
specified by one of the generators configurable parameters.
Each centroid, which represents a distinct class, is given a
random standard deviation and a multivariate center that is a
proper distance from all the other centroids centers. Not
ensuring a proper distance between centroids leads to
ambiguous results, because the resulting radial fields
associated with two or more centroids may overlap. In some
cases, the amount of overlap is considerable. The number of
variables or attributes assigned to the centroids centers is also
specified via a configurable parameter. A new data record is
generated by first randomly selecting one of the centroids.
Then a random offset with direction is created from the
chosen centroids center. The magnitude of the offset is
randomly drawn from a Gaussian distribution in combination
with the centroids standard deviation. This effectively creates
a normally distributed hypersphere of data records, with
distinct density, around the corresponding centroid [14].
The first series of experiments, with the synthetic RBF
generator, configured the generator to produce a data stream
comprising 200,000 data records with 5 classes and whose
records had five attributes. The MBT was set to 3.0 and the
sliding time window captured the entire data stream. On some
occasions, the result was as expected; i.e., five very dense
microclusters with radii ranging from 0.18 to 1.4 and whose
population of absorbed data records was rather evenly
distributed. On other occasions, the result was more than five
Vol. 2, No. 11, 2011
98 | P a g e
microclusters. However, on these occasions, there were
always five dense microclusters that had no overlap and
absorbed the vast majority of the data streams records. Of
those five, there was always one that had absorbed
substantially less data records than the other four. Using the
CQL, it was noted that there was a considerable amount of
overlap between this one microcluster and the other sparsely
populated extra microclusters. When the CluSandra
frameworks aggregator was run with an overlap factor of 1.0,
that one microcluster absorbed all the extra sparsely populated
microclusters. There was also a rather even distribution of
data records across all five microclusters. A second series of
experiments was executed that was identical to the first; the
one exception being that the generator was configured to
produce 7 classes, instead of 5. The results were consistent
with those of the previous series of experiments.
VI. CONCLUSIONS AND FUTURE WORK
This work presents a distributed framework and algorithm
for clustering evolving data streams. The Java-based
framework, which is named the CluSandra framework,
exhibits the following characteristics:
- It is entirely composed of proven open source
components that can be deployed on a variety of
commodity hardware and operating systems;
therefore, it is very economical to implement.

- It is leveraged by clustering processes to address the
severe time and space constraints presented by the
data stream. In other words, the functionality required
to address these constraints is offloaded from
clustering process to the framework, and it is the
framework that controls the data stream.
- It provides a distributed platform through which
ensembles of clustering processes can be seamlessly
distributed across many processors. This results in
high levels of reliability, scalability, and availability
for the clustering processes.
- It allows its hosted ensembles of clustering processes
to elastically scale up and down to meet the most
demanding dynamic data stream environments.

- It provides convenience classes or objects whose
purpose is to facilitate the implementation of
clustering algorithms and their subsequent
deployment onto the framework.
- It provides an effective mechanism through which the
hosted clustering processes can reliably and
efficiently store their byproduct of real-time statistical
summary information (i.e., microclusters) in a cluster
database.
- It provides a Cluster Query Language (CQL) that is
used to perform near-time or offline analytics against
the cluster database. The CQL offloads the offline
analytics from the clustering algorithm and provides a
mechanism for the implementation of a variety of
offline analytical processes. The CQL can also be
used by offline processes to monitor, in near-time,
clusters in the database and raise alerts whenever
clusters of a particular nature appear and/or disappear.
- It lays down the foundation for a data management
system whose focus is on clustering high speed and
evolving data streams.
The algorithm developed is named the CluSandra
algorithm and it is based upon concepts, structures, and
algorithms introduced in [10] and [5]. The algorithms
implementation serves as an example of a clustering process
that is designed to leverage the services and functionality
provided by the CluSandra framework. The algorithm is more
closely related to CluStream with its concept of viewing the
data stream as a changing process over time and its
functionality being divided between two operational phases:
real-time statistical data collection and offline analytical
processing. However, it deviates from CluStream primarily in
how it addresses these operational phases. Unlike CluStream,
it only performs the real-time statistical data collection, in the
form of microclustering, leaving the offline analytical
processing to end-users and/or processes that leverage the
CQL. This results in a simpler and higher performing
algorithm; primarily, because the agglomeration of
microclusters is not performed by the algorithm. It also
provides added flexibility to the end-user and/or offline
analytical processes, because it affords the opportunity to
analyze the collected data prior to agglomeration. Also, due to
its reliance on the framework, the CluSandra algorithm can
reliably and efficiently persist its microclusters to a cluster
database; this is addressed in neither [10] nor [5]. The end
result is the development of a clustering algorithm that
exhibits these characteristics: configurable, distributable,
elastically scalable, highly available and reliable, and simpler
to implement.
The following are topics for future work:
- The implementation of clustering algorithms designed
to address other data types and distributions, besides
numerical and Gaussian, and their deployment onto
the framework.
- The introduction of an integration framework, such as
[18], that allows for the quick implementation of
messaging design patterns meant to further control the
data stream and facilitate the implementation of multi-
staged data stream processing. For example, patterns
to address data processing steps such as cleansing,
standardization and transformation, and patterns used
for routing data records based on their content.
- The implementation of a graphical user interface that
leverages the CQL and provides a visual
representation of the data in the cluster database.
- Additional research and development on algorithms
that automate the calculation of an optimal MBT for
the target data stream.
Vol. 2, No. 11, 2011
99 | P a g e
REFERENCES
[1] C. Walls, Spring In Action, 2
nd
ed., Greenwick, CT: Manning
Publications Co., 2008.
[2] R. Pressman, Software Engineering A Practitioners Approach, 7
th
ed.,
New York: McGraw-Hill, 2009.
[3] P. Domingos, G. Hulten, Mining high-speed data streams, in
Knowledge Discovery and Data Mining, 2000, pp.71-80.
[4] P. Domingos, G. Hulten, L. Spencer Mining time-changing data
streams, in ACM KDD Conference, 2001, pp.97-106.
[5] C. Aggarwal, J. Han, J. Wang, P.S. Yu, A framework for clustering
evolving data streams, in Proceedings of the 29
th
VLDB Conference,
Berlin, Germany, 2003.
[6] J. Gama, Knowledge Discovery From Data Streams. Boca Raton, FL:
Chapman & Hall/CRC, 2010.
[7] J. Gama, M. Gaber, Learning From Data Streams, Processing
Techniques in Sensor Data, Berlin-Hiedelberg: Springer-Verlag, 2007.
[8] B. Bifet, R. Kirkby, R., Data Stream Mining, University of Waikato,
New Zealand: Centre for Open Software Innovation, 2009 .
[9] E. Hewitt, Cassandra, The Definitive Guide. Sebastopol, CA: OReilly
Media, Inc., 2010.
[10] T. Zhang, R. Ramakrishnan, M. Livny, BIRCH: an efficient data
clustering method for very large databases, in ACM SIGMOD, 1996,
pp. 103-114.
[11] G. Hebrail, Introduction to Data Stream Querying and Processing,
International Workshop on Data Stream Processing and Management,
Beijing: 2008.
[12] J. Han, M. Kamber, Data Mining Concepts and Techniques, 2
nd
ed., San
Francisco, CA: Morgan Kaufmann Publishers, 2006.
[13] B. Everitt, S. Landau, M. Leese, Cluster Analysis, 4
th
ed., London,
England: Arnold Publishers, 2001.
[14] B. Bifet, R. Kirkby, Massive Online Analysis Manual. University of
Waikato, New Zealand: Centre for Open Software Innovation, 2009.
[15] The Apache Cassandra Project. Available:
http://cassandra.apache.org, [Accessed Sept. 2011].
[16] Apache ActiveMQ. Available: http://activemq.apache.org, [Accessed
Sept. 2011].
[17] T. White, Hadoop: The Definitive Guide, Sebastopol, CA: OReilly
Media, Inc., 2010.
[18] Apache Camel. Available: http://camel.apache.org, [Accessed Sept.
2011].
[19] C. Plant, C. Bohm, Novel Trends In Clustering, Technische Universitat,
Munchen Germany, Ludwig Maximialns Universitat Munchen,
Germany, 2008.
[20] L. Kuncheva, Classifier ensembles for detecting concept change in
streaming data: overview and perspectives, in Proceedings of the
Second Workshop SUEMA, ECAI, Partas, Greece, July 2008, pp. 5-9.
[21] C. Aggarwal, J. Han, J. Wang, P.S. Yu, A framework for projected
clustering of high dimensional data streams, in Proceedings of the 30
th

VLDB Conference, Toronto, Canada, 2004, pp. 852-863.
AUTHORS PROFILES
Jose R. Fernandez received his MSc (2011) in Computer Science at the
University of West Florida and his BSc (Eng., 1983) in Computer and
Information Sciences at the University of Florida. He has over 25 years of
commercial software engineering experience and has held senior architecture
and management positions at NCR, AT&T, and BEA Systems. He has also
been instrumental in starting several companies whose focus was on the
development of Java Enterprise Edition (JEE) software products. He is
currently a senior consultant with special research interests in machine
learning and data mining.

Eman M. El-Sheikh is the Associate Dean for the College of Arts and
Sciences and an Associate Professor of Computer Science at the University of
West Florida. She received her PhD (2002) and MSc (1995) in Computer
Science from Michigan State University and BSc (1992) in Computer Science
from the American University in Cairo. Her research interests include
artificial intelligence-based techniques and tools for education, including the
development of intelligent tutoring systems and adaptive learning tools,
agent-based architectures, knowledge-based systems, machine learning, and
computer science education.

Vol. 2, No. 11, 2011
100 | P a g e
Clustering: Applied to Data Structuring and Retrieval

Ogechukwu N. Iloanusi
Department of Electronic Engineering
University of Nigeria, Nsukka
Enugu State, Nigeria
Charles C. Osuagwu
Department of Electronic Engineering
University of Nigeria, Nsukka
Enugu State, Nigeria

AbstractClustering is a very useful scheme for data structuring
and retrieval behuhcause it can handle large volumes of multi-
dimensional data and employs a very fast algorithm. Other forms
of data structuring techniques include hashing and binary tree
structures. However, clustering has the advantage of employing
little computational storage requirements and a fast speed
algorithm. In this paper, clustering, k-means clustering and the
approaches to effective clustering are extensively discussed.
Clustering was employed as a data grouping and retrieval
strategy in the filtering of fingerprints in the Fingerprint
Verification Competition 2000 database 4(a). An average
penetration of 7.41% obtained from the experiment shows clearly
that the clustering scheme is an effective retrieval strategy for the
filtering of fingerprints.
Keywords-component; Clustering; k-means; data retrieval;
indexing.
I. INTRODUCTION
A collection of datasets may be too large to handle and
work on hence may be better grouped according to some data
structure. Large datasets are encountered in filing systems in
digital libraries, access to and caching of data in databases and
search engines. Given the high volume of data there is need for
fast access and retrieval of required or relevant data. Several of
the existing data structures are hashing [1, 2, 3, 4, 5, 6], search
trees [7, 8], and clustering [9]. Hashing is a technique that
utilizes a hash function to convert large values into hash values
and maps similar large values to the same hash values or keys
in a hash table. Clustering is however a useful and efficient
data structuring technique because it can handle datasets that
are very large and at the same time n-dimensional (more than 2
dimensions) and similar datasets are assigned to the same
clusters [9]. A 2D or 3D point can be imagined and illustrated
however it will be difficult to imagine or illustrate a 9-
dimensional data. When datasets are clustered, the clusters can
be used rather than the individual datasets.
Clustering is a process of organizing a collection of data
into groups whose members are similar in some way [9, 10, 11,
12] According to Jain et al. [13] Cluster analysis is the
organization of a collection of patterns (usually represented as
a vector of measurements, or a point in a multidimensional
space) into clusters based on similarity. Similarity is
determined using a distance measure and objects are assigned
and belong to the same cluster if they are similar according to
some defined distance measure. Cluster analysis differs from
classification because in clustering the data are not labeled and
hence are naturally partitioned by the clustering algorithm
whereas in classification the data are labeled and partitioned
according to their labels. The former is hence an unsupervised
mode of data structuring while the later is supervised [13]. Jain
[14] identifies three main reasons while data clustering is used;
to understand the underlying structure of the data; to determine
degree of similarity amongst the data in their natural groupings
and to compress data by summarizing the data by cluster
groups.
Clustering has a vast application in the life sciences,
physical and social sciences and especially in the disciplines of
Engineering and Computer Science. Clustering is used for
pattern analysis, recognition and classification, data mining and
decision making in areas such as document retrieval, image
processing and statistical analysis and modeling [13].
Documents may be clustered for fast information access [15] or
retrieval [16]. Clustering is used in image processing to
segment images [17] as well as in marketing, biology,
psychiatry, geology, geography and archeology [13]. Figure 1
shows a general data clustering illustration. The data are
grouped in clusters. Each cluster has a collection of data that
are similar.

Figure 1. Data Clustering
A cluster is a group of similar datasets represented by an n-
dimensional value given by the cluster centroid. Clusters may
also be defined as high density regions separated by low
density regions in the feature space [13].
Every cluster is assumed to have a centroid, which is the
arithmetic mean of all data in that cluster. The mean is what is
common to data assigned to a cluster and creation of clusters
Similar data
Dissimilar data
Data
Clustered data
Vol. 2, No. 11, 2011
101 | P a g e
builds from the arithmetic mean. A similarity measure is used
for the assignment of patterns or features to clusters.
II. CLUSTER SIMILARITY MEASURES
Similarity is fundamental to the definition of a cluster hence
a measure for the similarity otherwise known as the distance
measure is essential. The dissimilarity or similarity between
points in the feature space is commonly calculated in cluster
analysis [13]. Some of the distance measures used are:
Euclidean distance
Manhattan distance
Chebyshev distance
Hamming distance
The distance metric is used for computing the distance
between two points and cluster centers. For the distance
measures explained in the following sections, two points, a and
b, are defined in an n-dimensional space as:
a = (w
0
, x
0
, y
0
z
0
) coordinates (1)
b = (w
1
, x
1
, y
1
z
1
) coordinates (2)
A. Euclidean distance
Euclidean distance is the distance between two points, a
and b, as the crow flies in an n-dimensional space.
( ) ((
) (3)
(
(4)

where n is the number of dimensions. The Euclidean
distance is the most commonly used metric because it is
appealing to use in an n-dimensional space and it works well
with isolated clusters [13].
B. Manhattan distance
In the Manhattan distance, the distance between two points
is the absolute difference of their coordinates.
( ) (
) (
) (5)
The difference between the Euclidean distance and the
Manhattan distance is that the Euclidean is a squared distance
while the Manhattan is not squared.
C. Chebyshev distance
In the Chebyshev distance metric the distance between two
points is the greatest of their differences along any coordinate
dimension [18]. This distance is named after Pafnuty
Chebyshev.
( ) (
) (6)
This is also known as the chessboard distance. In the
chessboard the length of side of a chess square may be assumed
as one unit. In this case the minimum number of moves needed
by a king to go from one chess square to another equals the
Chebyshev distance between the centers of the squares.
D. Hamming distance
The Hamming distance is a way of determining the
similarity of two strings of digits of equal lengths by measuring
the number of substitutions required to change a string into
another. It is the number of positions at which corresponding
digits in the two strings are different [19].
Given two strings a and b where
a = 0110110 and b = 1110011, the difference between the
two strings a and b, D(a,b), where

D(a,b) = 3, as the corresponding digits differ in three
places.
III. CLASSIFICATION OF CLUSTERING
ALGORITHMS
Clustering algorithms may be classified as:
Exclusive clustering
Overlapping clustering
Hierarchical clustering
A. Exclusive clustering
In exclusive clustering, data that belongs to a particular
cluster cannot belong to another cluster. An example is K-
means clustering.
B. Overlapping clustering
Data may belong to two or more clusters. Example of this
in fuzzy-c-means clustering.
C. Hierarchical clustering
In this case clusters are represented in tree from. Two close
clusters are derived from the top-level cluster. The hierarchy is
built by individual elements progressively merging into bigger
clusters.
Figure 2 shows the types of data clustering algorithms.
Jain [13] classifies clustering algorithms as hierarchical and
partitional. In hierarchical clustering each cluster arises from
and depends on the parent cluster. A typical partitional
clustering algorithm is the K-means algorithm.
IV. CLUSTER SIMILARITY MEASURES
K-means clustering algorithm was first proposed over 50
years ago [14] and is commonly preferred to other clustering
algorithms because of its ease of implementation and efficiency
in cluster analysis.
K-means clustering is a type of cluster analysis that
partitions n observations into k disjoint clusters, k<<n, such
that the number of clusters are much less than the number of
observations [18, 20]. The k-means algorithm partitions n
observations {O
i
| i=1, 2...n} into k number of clusters, {C
j
|
j=1,2...k}, as follows
{
} *
+ (7)
This is illustrated in Figure 3.
Vol. 2, No. 11, 2011
102 | P a g e

Figure 2. Types of Data Clustering Algorithms
Any of the distance metrics which include the Euclidean,
Manhattan, Chebyshev or Hamming may be used as the
distance measure for determining the similarity of the datasets,
though the Euclidean is most preferred and widely used [14].
The K-means algorithm basically follows these steps.
A similarity or distance measure is chosen and
used throughout.
K number of centroids are chosen.
The distance between each dataset from each of
the k centroids is determined.
Then a dataset is assigned to the centroid for
which it had the minimum distance.
All datasets are hence assigned to a particular
centroid. Figure 4 shows a very simple illustration
using 5 datasets and 2 clusters.
The arithmetic mean is recalculated for each of the
k centroids and the distance of each dataset from
the new means is recalculated for each of the k
centroids. This is the second iteration.
The datasets are reassigned again to the new k
centroids. In other words, a dataset assigned for
instance to centroid 2 in the first iteration may be
reassigned to centroid 1 in the second iteration.

Figure 3. K-means Clustering
The arithmetic means is recalculated and the data-
sets reassigned again.
This continues to i number of iterations, and the
iteration stops when there is no change in the
assignments between the ith iteration and the (i-
1)th iteration.
The last k centroids are the k clusters.
Exclusive
Overlapping
Hierarchical
Observation, O
i
;
i =1,2n
n observations
Clusters, C
j
;
j =1,2k
C
1

C
2

C
k

Vol. 2, No. 11, 2011
103 | P a g e

Figure 4. K-means clustering algorithm
V. APPROACHES TO EFFECTIVE CLUSTER ANALYSIS
Several challenges to cluster analysis include the
difficulties in choosing an appropriate clustering algorithm;
representing the data to be clustered; choosing a suitable
similarity measure; determining which data should be used and
choosing a suitable number of clusters that would yield
maximum success. A user is not faced with these problems in
hierarchical clustering analysis since all the datasets are related.
These problems arise when a dataset is to be classified into
unique clusters.
There are many partitional clustering algorithms and the
user may be faced with the dilemma of choosing an appropriate
algorithm. What guides in choosing an appropriate algorithm is
to know the purpose and goal of the clustering exercise and this
consequently would guide in representing that data to be
clustered. It is also necessary to know if the dataset has a
clustering tendency [18] and if it should be normalized. A data
set that does not have a clustering tendency should not be
clustered as it would yield invalid clusters. For example, if a
data set that has all similar data and consequently has no
variance is clustered it would result in invalid clusters. On the
other hand a data set with high variance has a clustering
tendency.
The choice of number of clusters may pose a problem
because the performance of the clustering algorithm is affected
by the number of clusters. It is usually difficult to determine the
best number of partitions that will give the best and valid
clustered groups. Some factors that may be considered while
choosing the number of clusters are the size of the data set and
the variance of the data in the data set. If the dataset is widely
varied such that the data set may need to be classified by many
groups then it may make sense to use more clusters.
In feature classification, the success of the cluster analysis
is largely dependent on the feature set. The clustering algorithm
would have a good performance and give compact, isolated and
valid clusters if the choice of features is good [18]. If for
instance a database of face images of a multi-racial group
comprising African, Chinese, Latin American, Indian and
European faces need to be clustered into five different groups,
the features that would be used would be such that the faces
can be effectively separated into five valid clusters. The
success of this task is clearly dependent on the features used for
the separation. The features making up the data set play a vital
role in clustering analysis.
A similarity measure is required for separating data into
clusters. The choice of the similarity measure is a challenging
problem because the valid clustering of the data also depends
on the similarity metric. The performance of the cluster
analysis varies according to the similarity metric used and
hence it may be difficult to determine the similarity metric that
would give the best performance. But this problem can be
overcome by having a good understanding of the data to be
clustered.
VI. CLUSTERING USED AS A FINGERPRINT INDEXING
RETRIEVAL STRATEGY
An indexing technique must include a retrieval strategy. A
retrieval strategy defines the method for which data within the
same class as the query or input data are retrieved. In
fingerprint indexing, the retrieval strategy ensures that
fingerprints with similar index codes [21] to that of the query
fingerprint are retrieved from the database of enrolled
fingerprints.
In this work, a modified Rosss partitional clustering
scheme [22] is used as a fingerprint retrieval strategy by
compressing the numerous fingerprint features into similar
groups of data and hence limiting the search for similar
fingerprints to only a few clusters that are identical to the
cluster of the query fingerprint. This requires the following
First the creation of an index space of k clusters for
the indexing using the k-means algorithm and the Euclidean
distance similarity measure.
Secondly the assignment of the features of the
fingerprints in the enrolled database to the k clusters.
Thirdly the determination of the clusters, c << k, that
have the features of the fingerprints similar to a query
fingerprint.
A query fingerprint should have a matching identity in a list
of fingerprints outputted by the indexing algorithm. This list is
otherwise known as the candidate list. The ratio of the
fingerprints in the candidate list to the database size gives the
penetration rate of a query fingerprint. The penetration rate is
the fraction of fingerprint identities, including the genuine
fingerprint, retrieved from the database upon presentation of an
input fingerprint. The penetration rate determined for a number
of tests, T, in a database of size, N, is [23].

()
1. Datasets to be clustered

2. Centroids a and c chosen

3. Distances of datasets from
centroids are determined

4. Datasets are assigned to
centroids (first iteration)

a
b
c
d e
a
b
c
d
e
a
b c
d
e
a
b c
d
e
Vol. 2, No. 11, 2011
104 | P a g e
Where {C
j
| j = 1, 2T} is the size of candidate list of the
fingerprints. The less the penetration rate the better the
performance of the algorithm.
In this newly created file, highlight all of the contents and
import your prepared text file. You are now ready to style your
paper; use
VII. EXPERIMENTS AND RESULTS
The Fingerprint Verification Competition (FVC)
2000 database 4(a) and FVC 2002 database 4(a)
were used for this experiment. Each database has
800 fingerprints from 100 subjects at 8 impressions
per subject.
The fingerprint features were extracted using the
minutiae quadruplets technique [24].
30 clusters were created in the index space using
fingerprints from FVC 2002 database 4(a).
The FVC 2000 database 4(a) was divided into two
equal groups Group A and B.
Group A had 400 fingerprints of the first four
impressions of a subject
Group B had 400 fingerprints of the last four
impressions of a subject.
The fingerprint features of group A were assigned
to the 30 clusters in the index space.
The fingerprints of group B were used to query the
index space to find a matching identity determined
by the penetration of the database. Every query
resulted in a penetration rate. Majority of the
queries had little penetration rates while some had
long penetration rates.
The penetration rates of the 400 query fingerprints used in
the experiments are shown in Table 1.
TABLE I. PENETRATION RATES OF 400 QUERY FINGERPRINTS IN AN
EVALUATION ON FVC 2000
No of tests
(T)
Penetration rates
(range)
Average
value fx
57 0.25 - 1.00 0.75 42.75
55 1.25 - 2.00 1.625 89.375
39 2.25 - 3.00 2.625 102.375
49 3.25 - 4.50 3.875 189.875
50 4.75 - 6.50 5.625 281.25
47 6.75 - 8.50 7.625 358.375
55 8.75 - 14.25 11.5 632.5
48 14.5 - 38.25 26.375 1266
T=400 fx=2962.5

The average penetration for the 400 query fingerprints is
obtained using Equation (8) and can also be determined from
Table 1 as:

(9)

Where f
x
is the product of the first and third columns in Table I
and T is the number of queries corresponding to the number of
tests in the experiment. There were 400 queries.
The retrieval of a candidate list for a query fingerprint takes
0.592ms.
VIII. COMPARISON WITH OTHER DATA STRUCTURING
TECHNIQUES
In [25], a binary tree based approach was used for matching
fingerprints. The work done on this paper is indexing.
However, the computational time for a fingerprint match using
the binary tree technique in [25] is compared with the
computational time for indexing a query fingerprint using the
clustering technique described in this paper in Table II.
TABLE II. COMPARISON OF COMPUTATIONAL TIMES OF THE BINARY
TREE AND CLUSTERING TECHNIQUES ON FVC 2002 DB1 SET A
Technique Database size Computational time
Binary tree [25] 800 fingerprints 34.8ms
Clustering (our
approach)
400 fingerprints 0.592ms
IX. CONCLUSION
In this paper, clustering was discussed extensively.
Experiments were conducted by employing a modified
clustering scheme as a retrieval strategy for filtering
fingerprints. The average penetration, 7.41%, is very small
showing clearly that the clustering algorithm employed is an
effective scheme for the filtering and retrieval of the candidate
fingerprints to a given query fingerprint.
REFERENCES
[1] H. J. Wolfson, Geometric Hashing: An Overview IEEE
Computational Science and Engineering, pp. 10 21, 1997.
[2] S. Danker, R. Ayers and R. P. Mislan, Hashing Techniques for Mobile
Device Forensics SMALL SCALE DIGITAL DEVICE FORENSICS
JOURNAL, vol. 3, no. 1, June 2009 ISSN:1941-6164
[3] D. Zhang and J. Wang, Self-Taught Hashing for Fast Similarity
Search SIGIR10, July 1923, 2010, Geneva, Switzerland
[4] H. Lim, J. Seo and Y. Jung, High speed IP address lookup architecture
using hashing IEEE Communication Letters, vol. 7, no. 10, pp. 502
504, ISSN: 1089-7798, October 2003
[5] D. Greene, M. Parnas and F. Yao, Multi-index hashing for
information retrieval Proceedings of the 35th Annual Symposium on
the Foundations of Computer Science, 1994, pp. 722 731
[6] X. Nie, J. Liu, J. Sun and W. Liu, Robust Video Hashing Based on
Double-Layer Embedding IEEE Signal Processing Letters, vol. 18, no.
5, pp. 307 310, May 2011
[7] B. Pfaff, Performance Analysis of BSTs in System Software
SIGMETRICS/Performance04, New York, NY, USA. June 1216,
2004,
Vol. 2, No. 11, 2011
105 | P a g e
[8] D. D. Sleator and R. E. Tarjan, Self-Adjusting Binary Search Trees
Journal of the Association for Computing Machinery. vol. 32, no. 3, July
1985, pp. 652-686.
[9] A.K. Baughman, S. Stockt, and A. Greenland, Large Scale Fingerprint
Mining. 2010 ACM 978-1-4503-0220-3.
[10] A. K. Jain, A. Topchy, M. Law, and J. Buhmann, "Landscape of
Clustering Algorithms", in Proceedings of the 17th International
Conference on Pattern Recognition, Cambridge UK, August 23-26, 2004
pp. I-260--I-263.
[11] F. Alberto and G. Sergio, "Solving Non-uniqueness in Agglomerative
Hierarchical Clustering Using Multidendrograms". Journal of
Classification, Vol.25, 2008, pp.4365.
[12] D. Apetrei, P. Postolache, N. Golovanov, M. Albu and G. Chicco,
Hierarchical Cluster Classification of Half Cycle Measurements in Low
Voltage Distribution Networks for Events Discrimination. International
Conference on Renewable Energies and Power Quality (ICREPQ09).
[13] A. Jain, M. N. Murty and P. Flynn, Data clustering: A review. ACM
Computing Surveys, Vol.3, no. 13, 1999, pp. 264323.
[14] A.K. Jain, Data Clustering: 50 Years Beyond K-Means. Pattern
Recognition Letters, Vol. 31, No. 8, 2010, pp. 651-666.
[15] M. Sahami, Using Machine Learning to Improve Information Access.
Ph.D. Thesis, Computer Science Department, Stanford University. 1998.
[16] S. Bhatia and J. Deogun, Conceputal clustering in information
retrieval. IEEE Transactions. Systems Man Cybernet. Vol. 28 (B),
1998, pp. 427436.
[17] A. K. Jain and P. Flynn, Image segmentation using clustering. in
Advances in Image Understanding. IEEE Computer Society Press, 1996,
pp. 6583.
[18] R.M.C.R. Souza and F.A.T. Carvalho, Dynamic clustering of interval
data based on adaptive Chebyshev distances. Electronics Letters, Vol
40, Issue 11, ISSN: 0013-5194, 2004, pp. 658 660.
[19] L. Gsieniec, J. Jansson and A. Lingas, Approximation algorithms for
Hamming clustering problems. ELSEVIER. Journal of Discrete
Algorithms, Vol. 2, 2004, pp. 289301.
[20] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman,
and A. Y. Wu, "An Efficient k-Means Clustering Algorithm: Analysis
and Implementation" IEEE Transaction on Pattern Analysis and
Machine Intelligence, Vol. 24, No. 7, July 2002. pp 881-892.
[21] D. Maltoni, D. Mario, A.K. Jain, and S. Prabhakar, Handbook of
Fingerprint Recognition. 2nd Ed. Springer-Verlag London Limited.
2009.
[22] A. Ross and R. Mukherjee, Augmenting Ridge Curves with Minutiae
Triplets for Fingerprint Indexing. In: Proc. of SPIE Conference on
Biometric Technology for Human Identification IV. Orlando, USA.
(April 2007).
[23] A. Gyaourova and A. Ross, A Novel Coding Scheme for Indexing
Fingerprint Patterns. In Proc. Joint IAPR International Workshop on
Structural, Syntactic, and Statistical Pattern Recognition. Springer-
Verlag, 2008, pp 755 -764.
[24] O. Iloanusi, A. Gyaourova, A. Ross, "Indexing Fingerprints Using
Minutiae Quadruplets," Proc. of IEEE Computer Society Workshop on
Biometrics at the Computer Vision and Pattern Recognition (CVPR)
Conference, (Colorado Springs, USA), 20-25 June 2011, pp 127 - 133.
[25] M. D Jain, N. S. Pradeep, C. Prakash and B. Raman, Binary tree based
linear time fingerprint matching, IEEE International Conference on
Image Processing (ICIP), 2006
AUTHORS PROFILE
Dr. Ogechukwu Iloanusi is a Lecturer at the Department of Electronic
Engineering, University of Nigeria, Nsukka. She holds a Ph.D in Digital
Electronics and Computer. Her research interests are biometric recognition
and algorithm development, web-based applications, micro-processor based
system design and e-learning..
Prof. Charles Osuagwu is a Professor of the Department of Electronic
Engineering, University of Nigeria, Nsukka. He has an M. Sc and Ph.D in
Engineering from the University of Southampton. His major research interests
are Digital Signal Processing, Microprocessor-based System Design, Software
Design, Computer Design, e-Security and Institutional Improvement.

Vol. 2, No. 11, 2011
106 | P a g e
Irrigation Fuzzy Controller Reduce Tomato Cracking

Federico Hahn
Irrigation Department
Universidad Autonoma Chapingo
Chapingo, Texcoco, Mxico

AbstractSunlight heats the greenhouse air temperature and
tomato cracking decreases marketable product up to 90%. A
shade screen reduced incoming radiation during warm and sunny
conditions to reduce tomato cracking. A fuzzy controller
managed greenhouse irrigation to reduce tomato cracking using
as variables solar radiation and substrate temperature. The
embedded controller presented 9 rules and three assumptions
that made it operate better. Signal peaks were removed and
control actions could take place ten minutes after irrigation.
Irrigation was increased during the peak hours from 12:00 to
15:00 h when it was required by the fuzzy controller; meanwhile
water containing the nutrient solution was removed during very
cloudy days with limited photosynthesis. After three continuous
cloudy days irrigation should be scheduled to avoid plant
nutrient problems. Substrate temperature in volcanic rock can be
used as real time irrigation sensor. Tomato cracking decreased to
29% using the fuzzy controller and canopy temperature never
exceeded 30C.
Keywords-component; Fuzzy controller; irrigation controller;
tomato cracking.
I. INTRODUCTION
Sunlight intensity control decrease irrigation needs as
plants and soil are kept cooler, increasing plant growth as
temperature drops off [1]. Cuticle cracking is a physiological
disorder that occurs at the beginning of the last phase of fruit
growth decreasing skin elasticity [2]. Tomato becomes
unmarketable for fresh consumption with losses ranging from
10% to 95% of the total fruit. Summer crack unmarketable
tomatoes are caused by high air temperature within the
greenhouse [3], so shading becomes an alternative, although
tomato yield decreases. Retractable roof shade houses are
structures covered with polypropylene, polyethylene, or
composite fabrics save up to 30% in energy [4]. Screen
systems reduce heat radiation losses at night, and decrease the
energy load on the greenhouse crop during sunny conditions.
Constant daily light integral achieved only through shading or
supplemental lighting [5], induces plant growth passing from
its initial transplanting state to its harvest state in 25 days.
Most of the irrigation systems uses ON/OFF controllers,
but optimal results are difficult to achieve due to varying time
delays and system parameters. A computer-based system
controlled irrigation and fertilization in greenhouses; sensors
provided feedback information for operating pumps and
solenoid valves [6]. Fuzzy theory interprets real uncertainties
and becomes ideal for nonlinear, time varying and hysteretic
system control. A fuzzy controller system saved water in
greenhouses and was cheap to implement [7]. A greenhouse
fuzzy climate controller required of 81 rules for proper
operation [8]. A simple fertigation fuzzy control presented
potential to save water and nutrients [9].
The following paper describes the design, implementation
of a fuzzy greenhouse controller which monitored radiation and
substrate temperature. Additional nutrients and water were
supplied to the greenhouse by an irrigation control to reduce
tomato cracking during the hot summer.
II. MATERIALS AND METHODS
The 80 x 40 m University greenhouse at Tlapeaxco,
Chapingo was used for the experiment, having natural cooling
through lateral and cenital openings. A black Raschel net
having a 50% of transmission was opened manually during the
first three months and closed during the summer when
tomatoes were produced. Three rows along the greenhouse
were transplanted on March 2th, 2009 in volcanic rock
substrate (dimensions: 40 x 40 x 10 cm) with tomato seedlings
(Lycopersicon esculentum var. Roma). A solenoid controlled
water and nutrient application in each row having tomato plants
spaced 30 cm; the irrigation treatments were: A, B and control.
Treatment A used the fuzzy controller to apply additional
irrigation during hot sunny days and removed water under
cloudy conditions. Treatment B added water with the fuzzy
controller under whichever weather condition. The last
treatment used the conventional controller and did not use any
kind of shading. Red tomatoes were harvested on the first of
June and one month later cracking was evaluated per treatment.
A fuzzy controller was selected for this application as
precise mathematical modeling of the controlled object is not
required becoming simpler to implement. Two photometric
sensors (model LI-210SB, Li-Cor Environmental Division,
USA) measured illumination providing 10 mV/100 klux being
cosine corrected up to 80 degree angle of solar incidence. The
sensors were calibrated against a standard lamp and presented a
sensitivity of 20 uA per 100 klux and a response time of 10
microseconds. One sensor was placed in the shading area, and
the other in the area without control. Radiation was acquired
every thirty seconds averaging ten continuous measurements.
Substrate temperature was measured with one
thermocouple fixed horizontally 8 cm from the pot top and
beneath the dripper. The J cooper constantan type
thermocouple was connected to a module (model TxRail 4-20
mA, Novus, Brasil) to provide a 4-20 mA signal finally
converted by means of a resistance to voltage. Platinum RTD
sensors (model RP502T22, Advanced Thermal Products, USA)
measured canopy temperature and was fixed beneath the
Vol. 2, No. 11, 2011
107 | P a g e
leaves. Six substrate and canopy temperature measurements
were acquired every 30 seconds from six plants selected
randomly. Temperature and radiation values were stored in a
datalogger (model CR1000, Campbell Scientific, USA). Daily
stored data were analyzed, comparing treatments behavior and
when high differences were noted a revision on the sensor
network and irrigation drippers was carried out.
A. Fuzzy membership functions and rules
Fuzzy evaluation methods process all the variables
according to predetermined weights and decrease the fuzziness
by using membership functions; therefore sensitivity is quite
high compared to other index evaluation techniques. The fuzzy
input membership functions determine the variables that are
going to be acquired to develop the control. It is desirable to
reduce the number of functions in order to minimize the
number of rules and the data contained in microcontroller
memory. Two input membership functions were used; the first
one employed the substrate temperature and the second one the
radiation, Fig. 1. The temperature function presented three
groups: COOL, FRESH and WARM temperature, being the
latter trapezoidal for values over 20C. The second function is
composed of three radiation groups named sunny (SUN),
cloudy (CLOU) and very cloudy (VC). The very cloudy group
of black and rainy sky presents the lowest radiation; a clean sky
radiates strongly and forms part of the sunny triangle.
Once the variables are acquired several rules establish the
actions that control the greenhouse tomato crop, Table I. The
two control actions are to indicate when to irrigate and to
increase irrigation and nutrient application. For example a
WARM temperature AND sunny radiation will irrigate the crop
(IRR) and add an additional 30% of water and 15% of nutrients
(ADNUT). In the case of a COOL temperature AND very
cloudy radiation (VC) the plants will not be irrigated (NON
IRR) and zero additional nutrition applied (ZNUT) due to poor
water and nutrient uptake. The AND condition was done with
the minimum algorithm. A real time clock controlled the fuzzy
rules during the 9:00-17:00 hr period. Every one and a half
hour starting at 9 AM drip irrigation was applied for seven
minutes. A total of seven periods were programmed per day,
but the fuzzy controller could apply five minutes more during
the irrigation periods of 12:00, 13:30 and 15:00, increasing
water application by 30%.
Sun COOL FRESH
5 10 16 19 28
Substrate temperature, C
WARM
VC Clou
10 25 45 70 100
Rad, klux
1
1

Figure 1. Temperature and radiation membership function
An embedded system used an ATM89C51 micro-controller
to control the solenoids to regulate water application per row.
TABLE I. FUZZY RULES TO CONTROL THE IRRIGATION SYSTEM
Radiation Nutrient and Water application

Cool Fresh Warm
VC ZNUT/NON IRR ZNUT/NON IRR ZNUT/NON IRR
SC ZNUT/ NON IRR ADNUT/ IRR ZNUT/ IRR
S ZNUT/ IRR ADNUT/ IRR ADNUT/ IRR
Substrate temperature values between 5 and 28C were
converted to a signal varying from 0 to 5 volts. The ADC804
converter acquired the temperature and radiation signals and
the digital output was sent to port 1. The real time clock (DS
1706) was connected to port 3. The fuzzy values were
transformed to its membership function using truth tables.
Irrigation control turned-on the pump and the solenoids,
according to the fuzzy rules used for adding water and
nutrients. The same fuzzy rules were carried out by treatments
A and B, but the microcontroller did not remove irrigation on
days with limited illumination in treatment B.
III. RESULTS AND DISCUSSION
Soil temperature inside the greenhouse nearby the substrate
was measured at 9:00, 12:00 and 15:00 h, Table II. At noon the
soil temperature covered with the black screen was 5.1C lower
than soil exposed directly to the sun. Air temperature measured
at a height of 3 m was 7.2C hotter than the soil temperature.
Inside greenhouses, leaves can be cooler than air decreasing
transpiration and increasing leaf condensation.

Figure 2. Substrate temperature affected by the shade screen.
Substrate temperature with and without shade becomes
alike at 1:00 PM; during the morning shaded substrate
temperature is higher, Fig. 2. Blue line discontinuities result
from irrigation. Temperature difference between canopy and
substrate was higher when no shading was present and
maximum values were encountered in sunny mornings. Many
peaks were encountered during unsteady climatic conditions,
causing multiple screen movement. Three assumptions were
considered for a better controller performance:

15
17
19
21
23
25
27
29
1
0
:
0
0
:
0
0
1
0
:
1
5
:
0
0
1
0
:
3
0
:
0
0
1
0
:
4
5
:
0
0
1
1
:
0
0
:
0
0
1
1
:
1
5
:
0
0
1
1
:
3
0
:
0
0
1
1
:
4
5
:
0
0
1
2
:
0
0
:
0
0
1
2
:
1
5
:
0
0
1
2
:
3
0
:
0
0
1
2
:
4
5
:
0
0
1
3
:
0
0
:
0
0
1
3
:
1
5
:
0
0
1
3
:
3
0
:
0
0
1
3
:
4
5
:
0
0
1
4
:
0
0
:
0
0
1
4
:
1
5
:
0
0
1
4
:
3
0
:
0
0
1
4
:
4
5
:
0
0
1
5
:
0
0
:
0
0
1
5
:
1
5
:
0
0
1
5
:
3
0
:
0
0
without shade
with shade T
e
m
p
e
r
a
t
u
r
e
,

C

Vol. 2, No. 11, 2011
108 | P a g e
1. At least three peaks were required in the same
direction during a period of 15 minutes
2. If one peak occurred in one direction and the next in
opposite direction they were cancelled.
3. After irrigation a control action could take place after
10 minutes.
TABLE II. SOIL TEMPERATURE MEASURED AT THREE DIFFERENT
HOURS IN A BLACK SHADED GREENHOUSE.

Soil temperature, C Light, Air temp,
Full sun 30% Black shade
klux C
09:00 a.m. 17.6 19.5
16
28
12:00 a.m. 38.2 33.1 70 40.3
03:00 p.m. 42 37.7 62 42

The three conditions used by the fuzzy controller under
operation made it perform better removing instantaneous
temperature and radiation peaks. The third rule was the most
useful as substrate temperature changes occurred just after
irrigation, affecting the substrate temperature signal.
The fuzzy controller was tested when step irrigation was
applied; canopy temperature changed after fifteen minutes,
Fig. 3. Substrate temperature decreased by 5C after the
twelve minute irrigation session (7 minutes of the normal
cycle + 5 minutes of the control) in the cloudy afternoon.
Although water is applied for twelve minutes, substrate
temperature variations will be noted during 20 minutes.
Substrate temperature difference between cloudy and sunny
days recorded every 5 min and averaged over each 60-min
interval were encountered at the moment of maximum
temperature [10]. The fuzzy controller takes ten data and after
processing one value will appear every 5 minutes. Substrate
daily temperature surpassed soil temperature and when it
exceeded 35C for more than 6 hours/day plant injury was
expected decreasing internodes spacing [11].
Maximum substrate temperature achieved throughout the
month was of 30C and reduced with irrigation as noted in
Fig. 3. If root water absorption is less than the rate of
transpiration loss, stomata will close, increasing plant and air
temperature; transpiration rate reduction in a shaded
greenhouse decreased total crop water uptake by 33% [12].

Figure 3. Radiation and temperature differences measured after irrigation.
Substrate temperature was considered a good indicator of
canopy temperature during all the day except during irrigation.
It can be used as a real time irrigation sensor for volcanic rock
substrate characterized by air spaces and low water retention.
Thirty hours were not irrigated due to electricity problems,
increasing the differential temperature between canopy and
substrate to 24C due to very little transpiration; Screens were
shot reducing the incoming radiation to the plants. Application
of continuous water in dark rainy days increased moisture to
80-90% so less irrigation was required and nutrients were
applied only twice a day. After three continuous cloudy days
irrigation should be scheduled for treatment A to avoid plant
nutrient problems.
Cuticle cracking appears under high illumination
conditions when temperature increases inside the greenhouse.
Daily tomato growth variations and tomato heating caused by
direct radiation cracked 52% of the tomatoes as noted on fruits
screen-uncovered; rapid tomato daily growth is avoided and
can withstand environmental and nutritional variations during
tomato growth. Cracking was reduced from 52% obtained
during the harvest of the first of June to 29% at the harvest of
the first of July using the fuzzy controller with treatment A. If
water lost by transpiration could not be replenished by the root
system, it will use water stored inside the fruit [13]. During
night, lower leaf transpiration and root pressure would
resupply water to the fruit. Factors such as volume and
irrigation frequency affect water uptake by the plants and
could be involved in the development of cracking [14].
Canopy temperature never exceeded 30C as additional
irrigation was added by the controller. Shading reduced the
radiation maintaining fresh the fruits; cracked harvested fruit
resulted from tomatoes having cracks at earlier stages.
Fruit weight increased when more irrigation was applied,
but in sunny days with high transpiration, water arrival to the
fruit decreased. Lack of irrigation once tomato weight exceeds
75 grams, affects its internal pressure and will start cracking
the fruits under direct heating. Average weight of the tomato
fruits with fuzzy control was 71 g being 6.4 g smaller than
tomatoes grown without control. Weights of tomatoes grown
under fuzzy control were more homogeneous as noted by the
standard deviation values of 0.77 and 0.69 for both treatments.
In treatment B, the fuzzy controlled supplied additional
irrigation and maintained the irrigation cycles during limited
illumination. As nutrient uptake by the plants is limited during
scarce illumination electrical conductivity increased resulting
in 31% of tomato cracking. The producer noted nutrient and
water losses during treatment B compared with treatment A
presenting both similar tomato average size. Plants without
shade presented losses of 52% by cracking, and although their
size was bigger and the yield higher. Different doses of
irrigation were applied when half of the days were sunny and
half cloudy being screen shading essential to reduce fruit
cracking.
Future work includes a movable controlled shade screen as
it was opened manually during this experiment; the screen
should be opened or closed automatically with a gear motor
driven by a photovoltaic system [15]. A pH and electrical
conductivity (EC) fuzzy system is required to control
Vol. 2, No. 11, 2011
109 | P a g e
fertigation more precisely, helping to observe cracking caused
by EC variations [9].
IV. CONCLUSION
This work concludes that tomato cracking can be reduced
by setting a screen; Cracking tomatoes harvested during July
were reduced from 52% to 29%. Raschel shading screens
reduced incoming radiation, fruit size and yield, but increased
marketable production and producer profit. The use of the
fuzzy controller regulated canopy temperature which never
exceeded 30C. Irrigation applied on sunny days helped to
maintain a constant fruit growth as well as a better canopy
temperature. Less water and fertilizers were used in treatment
A than in treatment B where irrigation was still applied under
scarce photosynthesis. Substrate temperature measurement
could be a new method for scanning real time irrigation in
volcanic rock substrates. The fuzzy controller becomes an
excellent option for plant growth management and healthy
tomato production but shade control should be included.
REFERENCES
[1] J. Tanny, S. Cohen, and A. Grava, Airflow and turbulence in a banana
screenhouse, Acta Horticulturae, vol. 719, pp. 623-630, 2006.
[2] J. C. Bakker, Russeting (cuticle cracking) in glasshouse tomatoes in
relation to fruit growth, J. Hort. Sci, vol. 63, pp. 459463, 1988.
[3] T. Wada, H. Ikeda, K. Matsushita, A. Kambara, H. Hirai, and K. Abe,
Effects of shading in summer on yield and quality of tomatoes grown
on a single-truss system, Journal Japan Society Horticultural Science,
vol. 75, pp. 51-58, 2006.
[4] H.F. Plaisier, and L. Svensson, Use of adapted energy screens in tomato
production with higher water vapour transmission, Acta Horticulturae,
vol. 691, pp. 583-588, 2005.
[5] L.D. Albright, A.J. Both, and A. Chiu, Controlling greenhouse light to a
consistent daily integral, Transanctions of the ASAE, vol. 43, pp. 421-
431, 2000.
[6] K. Kell, M. Beck, and F.W. Frenz, Automated ecological fertilization
and irrigation of soil grown crops in greenhouses with a computer
controlled system (KLIWADU), Acta Hortic., vol. 481, pp. 609-616,
1999.
[7] P. Javadi, A. Tabatabaee, M. Omid, R. Alimardani, and L. Naderloo,
Intelligent Control Based Fuzzy Logic for Automation of Greenhouse
Irrigation System and Evaluation in Relation to Conventional Systems,
Journal World Applied Sciences, vol. 6, pp. 16-23, 2009.
[8] F. Lafont, and J.-F Balmat, Optimized fuzzy control of a greenhouse,
Fuzzy Sets and Systems vol. 128, pp. 4759, 2002.
[9] D. Gmez, A. Lpez, G. Herrera, C. Fuentes, E. Rico, C. Olvera, D.
Alaniz, T. Mercado, and S. Verlinde, Fuzzy irrigation greenhouse
control system based on a field programmable gate array, Afr. J. Agric.
Res., Vol. 6, pp. 2544-2557, 2011.
[10] S. L.Warren, and T.E. Bilderback, Timing of low pressure irrigation
affects plant growth and water utilization efficiency, Journal
Environmental Horticulture, vol. 20, pp. 184188, 2002.
[11] F. Giuffrida, Temperature of substrates in relation to through
characteristics, Acta Horticulturae, vol. 559, pp. 647-654, 2001.
[12] P. Lorenzo, M.L. Garca, M.C. Snchez-Guerrero, E. Medrano, I.
Caparrs, and M. Gimnez, Influence of Mobile Shading on Yield,
Crop Transpiration and Water Use Efficiency, Acta Horticulturae, vol.
719, pp. 471-478, 2006.
[13] F. Jobin-Lawler, K. Simard, A. Gosselin, and A.P. Papadopoulos, The
influence of solar radiation and Boron-Calcium fruit application on
cuticle cracking of a winter tomato crop grown under supplemental
lighting, Acta Horticulturae, vol. 580, pp. 235-239, 2002.
[14] M. Dorais, A.P. Papadopoulos, and A. Gosselin, Influence of electrical
conductivity management on greenhouse tomato fruit yield and fruit
quality, Agronomie, vol. 21, pp. 367-383, 2001.
[15] F. Hahn, Fuzzy controller decreases tomato cracking in greenhouses,
Comput. Electron. Agric., vol. 77, pp. 21-27, 2011.
AUTHORS PROFILE
Dr. Federico Hahn
Professor and researcher of biosystems and
irrigation technology at the Universidad de Chapingo, Mexico.
Dr. Hahn develops control equipment for farmers and
producers. His embedded systems are working around Mexico
and other countries. His main areas of expertise are sensors,
instrumentation and automation.

Vol. 2, No. 11, 2011
110 | P a g e
Plethora of Cyber Forensics

N.Sridhar
1

Research Scholar, Dept.of CS&SE
neralla_sridhar@yahoo.com

Dr.D.Lalitha Bhaskari
2

Associate Professor, Dept.of CS&SE
lalithabhaskari@yahoo.co.in
Dr.P.S.Avadhani
3

Professor, Dept.of CS&SE
psavadhani@yahoo.com

Abstract As threats against digital assets have risen and there is
necessitate exposing and eliminating hidden risks and threats.
The ability of exposing is called cyber forensics. Cyber
Penetrators have adopted more sophistical tools and tactics that
endanger the operations of the global phenomena. These
attackers are also using anti-forensic techniques to hide evidence
of a cyber crime. Cyber forensics tools must increase its
toughness and counteract these advanced persistent threats. This
paper focuses on briefing of Cyber forensics, various phases of
cyber forensics, handy tools and new research trends and issues
in this fascinated area.
Keywords- Cyber Forensics; digital evidence; forensics tools; cyber
crimes.
I. INTRODUCTION
As Internet technologies proliferate into everyday life, we
come close to realizing new and existing online opportunities.
One such opportunity is in Cyber forensics, unique process of
identifying, preserving, analyzing and presenting digital
evidence in a manner that is legally accepted. The American
Heritage Dictionary defines forensics as relating to the use of
science or technology in the investigation and establishment of
facts or evidence in a court of law [1].
Computer forensics involves the identification,
documentation, and interpretation of computer media for using
them as evidence and/or to rebuild the crime scenario [2].
According to [3] computer forensics defined as the process of
identifying, collecting, preserving, analyzing and presenting
the computer-related evidence in a manner that is legally
acceptable by court.
More recently, computer forensics branched into several
overlapping areas, generating various terms [4] such as, digital
forensics, data forensics, system forensics, network forensics,
email forensics, cyber forensics, forensics analysis, enterprise
forensics, proactive forensics etc., as shown in figure-1.
Digital forensics is the investigation of what happened and
how. System forensics is performed on standalone machines.
Network forensics involves the collection and analysis of
network events in order to discover the sources of security
attacks. The same process applied on Web is also known as
Web forensics. Data forensics major focuses on analysis of
volatile and non-volatile data. Proactive forensics is an
ongoing forensics and there is an opportunity to actively, and
regularly collect potential evidence in an ongoing basis. Email
forensics deals with one or more e-mails as evidence in
forensic investigation. Enterprise forensics is named in the
context of enterprise; it is primarily concerned with incident
response and recovery with little concern about evidence.
Cyber forensics focuses on real-time, on-line evidence
gathering.
Forensics analysis deals with identification, extraction and
reporting on data obtained from a computer system.

Figure 1: Various types of Computer Forensics
A. Cyber Crimes
In the 2008 CSI Computer Crime and Security Survey, it
was noted that there is an average loss of $500,000 with
corporations experiencing financial fraud (related to
computing) and an extra average of $350,000 losses at
companies that experienced bot attacks [5]. As per [6], the
collecting digital devices in a forensically sound manner is
becoming more critical since 80% of all cases have some sort
of digital evidence involved in them.
With increased use of Internet in homes and offices, there
has been a proliferation of cyber-related crimes and these
crimes investigation is a tedious task. Cybercrime is typically
described as any criminal act dealing with computers or
computer Networks [7]. Cybercrimes can be classified into
three groups [2]; Crimes directed against computer, crimes
where the computer contains evidence, and crimes where the
computer is used to commit the crime. Other names of
cybercrime are e-crime, computer crime or Internet crime.
Cybercrimes spread across the world with various names like
worms, logic bombs, botnets, data diddling, mail bombing,
phishing, rootkits, salami theft, spoofing, zombie, time bomb,
Trojan horse etc.
Using the Internet, a person sitting in a Net cafe of a
remote location can attack a computer resource in USA using
Vol. 2, No. 11, 2011
111 | P a g e
a computer situated in Britain as a launch pad for his attack.
Challenges behind these situations are both technological and
jurisdictional. Confidentiality, integrity and availability are the
cardinal pillars of cyber security and they should not be
compromised in any manner [2]. Attackers also begin using
anti-forensic techniques to hide evidence of a cybercrime.
They may hide folders, rename files, delete logs, or change,
edit or modify file data [7]. To combat these kinds of crimes,
Indian Government established Cyber Forensics Laboratory in
November, 2003.
B. Overview of Cyber Forensics
Cyber forensics becoming as a source of investigation
because human expert witnesses are important since courts
will not recognize software tools such as Encase, Pasco,
Ethereal as an expert witness [8]. Cyber forensics is useful for
many professionals like military, private sector and industry,
academia, and law. These areas have many needs including
data protection, data acquisition, imaging, extraction,
interrogation, normalization, analysis, and reporting.
It is important for all professionals working in the
emerging field of cyber forensics to have a working and
functioning lexicon of terms like bookmarks, cookies, webhit
etc., that are uniformly applied throughout the profession and
industry. Cyber forensics international guidelines, related key
terms and tools are focused in the cyber forensics field manual
[7].
The objective of Cyber forensics is to identify digital
evidence for an investigation with the scientific method to
draw conclusions. Examples of investigations that use cyber
forensics include unlawful use of computers, child
pornography, and cyber terrorism.
The area of cyber forensics has become prominent field of
research because:
1) Forensics systems allow the administrator to diagnose
errors
2) Intrusion detection systems are necessary in avoiding
cyber crimes
3) Change detection can be possible with proactive
forensics
Cyber forensics can be used for two benefits [9]:
1) To investigate allegations of digital malfeasance
2) To perform cause analysis
II. PHASES OF CYBER FORENSICS
Cyber forensics has four distinct phases: incident
identification, acquisition of evidence, analysis of evidence,
and reporting with storage of evidence [10]. Figure 2 shows
various phases of cyber forensics process and each phase
responsibility. The identification phase mainly deals with
incident identification, evidence collection and checking of the
evidence. The acquisition phase saves the state of a computer
system that can be further analyzed. The analysis phase
collects the acquired data and examines it to find the pieces of
evidences. The reporting phase comprises of documentation
and evidence retention.

Figure 2: Phases of Cyber Forensics
A. Identification Phase
The identification phase is the process of identifying
evidence material and its probable location. This phase is
unlike a traditional crime scene it processes the incident scene
and documents every step of the way. Evidence should be
handled properly. Basic requirement in evidence collection is
evidence must be presented without alteration. This
requirement applies to all phases of forensics analysis. At the
time of evidence collection, there is a need of thorough check
of system logs, time stamps and security monitors.
Once evidence collected, it is necessary to account for its
whereabouts. Investigators would need detailed forensics to
establish a chain of custody, the documentation of the
possession of evidence. Chain of custody is a vital part of
computer forensics and the legal system [11] and the goal is to
protect the integrity of evidence, so evidence should be
physically secured in a safe place along with a detailed log.
Figure 3 shows the evidence and chain of custody which is
useful during incident investigation. Handling specific type of
incidents like Denial of Service, Malicious Code,
Unauthorized access etc are described in computer security
incident handling guide [12].

Figure 3: Evidence Form and Chain of Custody
B. Acquisition Phase
The acquisition phase saves the state of evidence that can
be further analyzed. The goal of this phase is to save all
digital values. Here, a copy of hard disk is created, which is
commonly called as an image. Different methods of acquiring
Vol. 2, No. 11, 2011
112 | P a g e
data and their relative advantages and disadvantages are
described in [13]. As per law enforcement community, there
are three types of commonly accepted forensics acquisition:
mirror image, forensics duplication and live acquisition.
Mirror images, bit-for-bit copy, involve the backups of
entire hard disk. Creation of mirror image is simple in theory,
but its accuracy must meet evidence standards. The purpose of
having mirror image is evidence available in the case of the
original system need to be restarted for further analysis.
A forensic duplicate, sector-by-sector, is an advanced
method that makes a copy of every bit without leaving any
single bit of evidence. The resultant may be single large file
and must be an exact representation of the original drive at bit-
stream level. This method is most common type of acquisition
because it creates a forensic image of the e-evidence and it
also contains file slack. In case of the small file overwrites a
larger file, surplus bytes are available in the file slack. The
forensic duplication process can be done with the help of tools
like Forensic Tool Kit (FTK) imager, UNIX dd command, or
Encase. Access Datas FTK is one of the powerful tools
available and one of the promising features is the ability to
identify steganography and practice of camouflaging data in
plain sight.
It is often desirable to capture volatile information, which
is stored in RAM; it cannot be collected after the system has
been powered down. This information may not be recorded in
a file system or image backup, and it may hold clues related to
attacker. All currently running processes, open sockets,
currently logged users, recent connections etc, are available in
volatile information.
Generally, intruder takes steps to avoid detection. Trojans,
keyloggers, worms etc., are installed in subtle places. One of
such things to be considered in the acquisition process is
rootkits, automated packages that create backdoors. An
Intruders/hackers use rootkits to remove log files and other
information to hide the presence of intruder. Mobile phones
are become tools for cybercrimes, mobile phone evidence
acquisition testing process are discussed in [14].
C. Analysis Phase
Forensic analysis is the process of understanding, re-
creating, and analyzing arbitrary events that have gathered
from digital sources [15]. The analysis phase collects the
acquired data and examines it to find the pieces of evidences.
This phase also identify that the system was tampered or not to
avoid identification. Analysis phase examines all the evidence
collected during collection and acquisition phases. There are
three types of examinations can be applied for the forensics
analysis; limited, partial or full examination.
Limited examination covers the data areas that are
specified by legal documents or based on interviews. This
examination process is the least time consuming and most
common type. Partial examination deals with prominent areas.
Key areas like log files, registry, cookies, e-mail folders and
user directories etc., are examined in this case of partial
examination. This partial examination is based on general
search criteria which are developed by forensic experts. Most
time consuming and less frequent examination process are full
examination. This requires the examiner to look each, and
every possible bit of data to find the root causes of the
incident. File slack inspection is done in this examination.
Some of tools used in the analysis phase are Coroner,
Encase, FTK. The Coroner toolkit run under UNIX and
EnCase is a toolkit that runs under Windows [7]. EnCase has
the ability to process larger amounts and allow the user to use
predefined scripts to pull information from the data being
processed. FTK contains a variety of separate tools (text
indexing, NAT recovery, data extraction, file filtering and
email recovery etc.) to assist in the examination.
D. Reporting Phase
The reporting phase comprises of documentation and
evidence retention. The scientific method used in this phase is
to draw conclusions based on the gathered evidence. This
phase is mainly based on the Cyber laws and presents the
conclusions for corresponding evidence from the
investigation. There is a need of good policy for how long
evidence from an incident should be retention. Factors to be
considered in this process are prosecution, data retention and
cost [12]. To meet the retention requirements there is a need of
maintaining log archival [16]. The archived logs must be
protected to maintain confidentiality and integrity of logs.
E. Forensics Methodology
The International Association of Computer Investigative
Specialists (IACIS) has developed a forensic methodology
which can be summarized as follows:
Protect the Crime Scene, power shutdown for the
computer and document the hardware configuration and
transport the computer system to a secure location
Bit Stream backup of digital media, use hash algorithms
to authenticate data on all storage devices and document
the system date and time
Search keywords and check file space management (swap
file, file slack evaluation, unallocated space)
Evaluate program functionality, document findings/results
and retain Copies of software
III. CYBER FORENSICS TOOLS
The main objective of cyber forensics tools is to extract
digital evidence which can be admissible in court of law.
Electronic evidence (e-evidence, for short) is playing a vital
role in cybercrimes. Computer forensics tools used to find
skeletons in digital media. To reduce the effect of anti-
forensics tools the Investigator is likely to have the tools and
knowledge required to counter the use of anti-forensics
techniques [17]. Sometimes collection of digital evidence is
straightforward because intruders post information about
themselves from Facebook, Orkut, Twitter, MySpace and chat
about their illegal activities. A subpoena, rather than special
forensics tools, required obtain this information; these e-mails
or chats from social networks can be admissible as evidence.
A snapshot of the state of the art of forensic software tools for
mobiles given in [18]. The process model for cellular phone
tool testing had shown in [14]. Various cyber forensics tools
and their description are provided in [7] some of them are:
Vol. 2, No. 11, 2011
113 | P a g e
1. The Coroners Toolkit (TCT), is an open source set of
forensic tools designed to conduct investigation UNIX
systems.
2. Encase is the industry standard software used by law
enforcement
3. The Forensic Toolkit (FTK) is very powerful tool but not
simple to use.
4. I2Analyst is a different type of analysis tool, It is a visual
investigative analysis software.
5. LogLogics LX 2000 is powerful and distributed log
analysis tool.
6. NetWitness and security intelligence, are network traffic
security analyzer tools.
7. ProDiscover Incident Response (IR) is a complete IT
forensic tool that can access computers over the network
to study the network behavior
8. The Sleuth Kit is one of network forensics tools used to
find file instances in an NTFS file
IV. CURRENT RESEARCH
Cyber Forensics is sizzling topic of the current trends.
Many researchers started doing intensive research in this
current area. New directions in this field include authorship
analysis, digital evidence collection and forensics
investigation process, proactive forensics, intrusion detection
systems with the help of honeypots, building evidence graphs,
identifying usage of mobile phones in cybercrimes and hash
function for preserving the integrity of evidence. The complete
picture of Cyber Forensics in the form of Cyber forensics
ontology which can be helpful for studying cyber forensics is
given in [19]. Proactive forensics helps in the creation of
preventive intelligence and threat monitoring besides post
incident investigations.
Advantages and disadvantages of intercepting wireless
network traffic as a means of locating potential evidence
sources during evidence seizure are listed in [20]. Also in the
same work the advantages and disadvantages of impairing
communications to or from 802.11-based wireless networks
during forensic seizure were discussed. High speed bitwise
search model for large-scale digital forensic investigations
using pattern matching board to search for string and complex
regular expressions discussed in [21].
Various methods on how the evidential value of digital
timestamps can be enhanced by taking a hypothesis based
approach to the investigation of digital timestamp in his thesis
work are proposed in the thesis [17]. Analysis of Instant
Messaging in terms of computer forensics and intrusion
detection is unexplored until now. Authorship classification
used for forensics analysis or masquerade detection [22].
Creation of mobile software that runs on a mobile device and
the goal is to aid crime scene personnel in the collection of
digital devices during the course of an investigation is
proposed by [23].
V. CONCLUSION
Cyber forensics is an emerging field in the 21
st
century.
Detailed study of the field of cyber forensics is given in this
paper. When analyzing cyber forensics, the process of doing
so is different from the traditional forensics. In this paper, we
described various computer forensics related definitions and
phases cyber forensics and forensics methodology.
Various phases of the Cyber forensics have been discussed
and each phase explored with their respective tools. Moreover,
we mentioned different tools that are utilized in cyber
forensics. Finally, we had shown the current research trends
in this new era of cyber forensics; it still evolves and will
remain a hot topic as long as there are ways to threaten data
security.
REFERENCES
[1] Kruse W.G, and Heiser J.G, Computer Forensics Incident Response
Essentials, 2002, Addison Wesley Pearson Education, Boston
[2] Ibrahim M. Baggili, Richard Mislan, Marcus Rogers, Mobile Phone
Forensics Tool Testing: A Database Driven Approach, International
Journal of Digital Evidence Fall 2007, Volume 6, Issue 2
[3] Caloyannides, Michael A, Computer Forensics and Privacy. Artech
House, Inc. 2001.
[4] Deepak Singh Tomar, Nikhil Kumar Singh, Bhopal Nath Roy, An
approach to understand the end user behavior through log analysis,
International Journal of Computer Applications (0975 8887) Volume
5 No.11, August 2010
[5] Svein Yngvar Willassen, Methods for Enhancement of Timestamp
Evidence in Digital Investigations, Doctoral thesis at NTNU, 2008: 19
[6] Wayne Jansen, Rick Ayers, Forensic Software Tools for Cell Phone
Subscriber Identity Modules, Conference on Digital Forensics, Security
and Law, 2006
[7] Ashley Brinson, Abigail Robinson, Marcus Rogers, A cyber forensics
ontology: Creating a new approach to studying cyber forensics, Digital
Instigation, Elsevier, 2006
[8] Benjamin Turnbull, Jill Slay, Wireless Forensic Analysis Tools for use in
the Electronic Evidence Collection, IEEE Proceedings of the 40th
Annual Hawaii International Conference on System Sciences-2007
(HICSS07)
[9] Hyungkeun Jee, Jooyoung Lee, and Dowon Hong, High Speed Bitwise
Search for Digital Forensic System, Proceedings of world academy of
science engineering and technology, volume 26, 2007.
[10] Angela Orebaugh and Jeremy Allnutt, Classification of Instant
Messaging Communications for Forensics Analysis, The International
Journal of Forensics Computer Science, 2009 (1), 22-28
[11] Ibrahim Baggili, Generating System Requirements for a Mobile Digital
Device Collection System, European and Mediterranean Conference on
Information Systems 2010, Abudhabi, UAE
AUTHORS PROFILE
N.Sridhar is a research scholar in Andhra University under
the supervision of Prof.P.S.Avadhani and Dr.D.Lalitha
Bhaskari. He received his M.Tech (IT) from Andhra
University and presently working as Associate Professor in
IT Department of GMRIT. He is a Life Member of ISTE.
His research areas include Network Securiy, Cryptography,
Cyber Forensics and Web Security.

Mrs. Dr. D. Lalitha Bhaskari is an Associate Professor in
the department of Computer Science and Engineering of
Andhra University. She is guiding more than 8 Ph. D
Scholars from various institutes. Her areas of interest
include Theory of computation, Data Security, Image
Processing, Data communications, Pattern Recognition.
Apart from her regular academic activities she holds prestigious
responsibilities like Associate Member in the Institute of Engineers, Member
in IEEE, Associate Member in the Pentagram Research Foundation,
Hyderabad, India.
Vol. 2, No. 11, 2011
114 | P a g e
Dr. P. S. Avadhani is a Professor in the department of
computer Science and Engineering of Andhra University.
He has guided 7 Ph. D students, 3 students already
submitted and right now he is guiding 12 Ph. D Scholars
from various institutes. He has guided more than 100
M.Tech. Projects. He received many honors and he has
been the member for many expert committees, member of Board of Studies
for various universities, Resource person for various organizations. He has co-
authored 4 books. He is a Life Member in CSI, AMTI, ISIAM, ISTE, YHAI
and in the International Society on Education Technology. He is also a
Member of IEEE, and a Member in AICTE.

Vol. 2, No. 11, 2011
115 | P a g e
A Fuzzy Similarity Based Concept Mining Model for
Text Classification
Text Document Categorization Based on Fuzzy Similarity Analyzer and Support Vector
Machine Classifier
Shalini Puri
M. Tech. Student
BIT, Mesra
India

AbstractText Classification is a challenging and a red hot field
in the current scenario and has great importance in text
categorization applications. A lot of research work has been done
in this field but there is a need to categorize a collection of text
documents into mutually exclusive categories by extracting the
concepts or features using supervised learning paradigm and
different classification algorithms. In this paper, a new Fuzzy
Similarity Based Concept Mining Model (FSCMM) is proposed
to classify a set of text documents into pre - defined Category
Groups (CG) by providing them training and preparing on the
sentence, document and integrated corpora levels along with
feature reduction, ambiguity removal on each level to achieve
high system performance. Fuzzy Feature Category Similarity
Analyzer (FFCSA) is used to analyze each extracted feature of
Integrated Corpora Feature Vector (ICFV) with the
corresponding categories or classes. This model uses Support
Vector Machine Classifier (SVMC) to classify correctly the
training data patterns into two groups; i. e., + 1 and 1, thereby
producing accurate and correct results. The proposed model
works efficiently and effectively with great performance and high
- accuracy results.
Keywords-Text Classification; Natural Language Processing;
Feature Extraction; Concept Mining; Fuzzy Similarity Analyzer;
Dimensionality Reduction; Sentence Level; Document Level;
Integrated Corpora Level Processing.
I. INTRODUCTION
From the long time, the discipline of Artificial Intelligence
(AI) is growing up on the map of science with psychology and
computer science. It is an area of study that embeds the
computational techniques and methodologies of intelligence,
learning and knowledge [1] to perform complex tasks with
great performance and high accuracy. This field is fascinating
because of its complementarities of art and science. It
contributes to increase the understanding of reasoning, learning
and perception. Natural Language Processing (NLP) is the
heart of AI and has text classification as an important problem
area to process different textual data and documents by finding
out their grammatical syntax and semantics and representing
them in the fully structured form [1] [2]. AI provides many
learning methods and paradigms to represent, interpret and
acquire domain knowledge to further help other documents in
learning.
Text Mining (TM) is a new, challenging and multi-
disciplinary area, which includes spheres of knowledge like
Computing, Statistics, Predictive, Linguistics and Cognitive
Science [3] [4]. TM has been applied in a variety of concerns
and applications. Some applications are summary creation,
clustering, language identification [5], term extraction [6] and
categorization [5] [6], electronic mail management, document
management, and market research with an investigation [3] [4]
[5].
TM consists of extracting regularities, patterns, and
categorizing text in large volume of texts written in a natural
language; therefore, NLP is used to process such text by
segmenting it into its specific and constituent parts for further
processing [2]. Text segmentation is also an important concern
of TM. Many researches go on for the work of text
classification [3] [4] [6] [7] [8] just for English language text.
Text classification is performed on the textual document sets
written in English language, one of the European Language,
where words can be simply separated out using many
delimiters like comma, space, full stop, etc. Most of the
developed techniques work efficiently with European
languages where the words, the unit of representation, can be
clearly determined by simple tokenization techniques. Such
text is referred as segmented text [10]. It does not always
happen. There are some Asian Languages in which textual
document does not follow word separation schemes and
techniques. These languages contain un-segmented text. There
are so many other languages in the world like Chinese and Thai
Languages in which there is no delimiter to separate out the
words. These languages are written as a sequence of characters
without explicit word boundary delimiters [10]. So, they use
different mechanisms for segmentation and categorization.
Here, the proposed effort is to work only on the English text
documents.
Therefore, TM categorization is used to analyze and
comprise of large volume of non - structured textual
documents. Its purpose is to identify the main concepts in a text
document and classifying it to one or more pre-defined
categories [3]-[12]. NLP plays an important and vital role to
convert unstructured text [4] [5] [6] into the structured one by
performing a number of text pre-processing steps. This
processing results into the extraction of specific and exclusive
Vol. 2, No. 11, 2011
116 | P a g e
concepts or features as words. These features help in
categorizing text documents into classified groups.
Section II discusses the thematic background and related
research work done on the concept mining, feature extraction,
and similarity measure. In section III, the proposed model and
methodology is discussed in detail. It discusses Text Learning
Phase which includes Text Document Training Processor
(TDTP), Pseudo Thesaurus, Class Feeder (CF) and Fuzzy
Feature Category Similarity Analyzer (FFCSA). Next, Support
Vector Machine Classifier (SVMC) and Text Classification
Phase are discussed. Finally, section IV concludes the paper
with the suggested future work and scope of the paper.
II. THEMATIC BACKGROUND AND RELATED WORK
Supervised learning techniques and methodologies for
automated text document categorization into known and
predefined classes have received much attention in recent
years. There are some reasons behind it. Firstly, in the
unsupervised learning methods, the document collections have
neither predefined classes nor labeled documents availability.
Furthermore, there is a big motivation to uncover hidden
category structure in large corpora. Therefore, text
classification algorithms are booming up for word based
representation of text documents and for text categorization.
A. Concept Mining
Concept Mining is used to search or extract the concepts
embedded in the text document. These concepts can be either
words or phrases and are totally dependent on the semantic
structure of the sentence. When a new text document is
introduced to the system, the concept mining can detect a
concept match from this document to all the previously
processed documents in the data set by scanning the new
document and extracting the matching concepts [5]. In this
way, the similarity measure is used for concept analysis on the
sentence, document, and corpus levels.
These concepts are originally extracted by the semantic role
labeler [5] and analyzed with respect to the sentence,
document, and corpus levels. Thus, the matching among these
concepts is less likely to be found in non - related documents. If
these concepts show matching in unrelated documents, then
they produce errors in terms of noise. Therefore, when text
document similarity is calculated, the concepts become
insensitive to noise.
B. Feature Extraction
In text classification, the dimensionality of the feature
vector is usually huge. Even more, there is the problem of
Curse of Dimensionality, in which the large collection of
features takes very much dimension in terms of execution time
and storage requirements. This is considered as one of the
problems of Vector Space Model (VSM) where all the features
are represented as a vector of n - dimensional data. Here, n
represents the total number of features of the document. This
features set is huge and high dimensional.
There are two popular methods for feature reduction:
Feature Selection and Feature Extraction. In feature selection
methods, a subset of the original feature set is obtained to make
the new feature set, which is further used for the text
classification tasks with the use of Information Gain [5]. In
feature extraction methods, the original feature set is converted
into a different reduced feature set by a projecting process. So,
the number of features is reduced and overall system
performance is improved [6].
Feature extraction approaches are more effective than
feature selection techniques but are more computationally
expensive. Therefore, development of scalable and efficient
feature extraction algorithms is highly demanded to deal with
high-dimensional document feature sets. Both feature reduction
approaches are applied before document classification tasks are
performed.
C. Similarity Measure
In recent years, fuzzy logic [1] [2] [4] [6]-[15] has become
an upcoming and demanding field of text classification. It has
its strong base of calculating membership degree, fuzzy
relations, fuzzy association, fuzzy production rules, fuzzy k-
means, fuzzy c-means and many more concerns. As such, a
great research work has been done on the fuzzy similarity and
its classifiers for text categorization.
The categorizer based on fuzzy similarity methodology is
used to create categories with a basis on the similarity of
textual terms [4]. It improves the issues of linguistic
ambiguities present in the classification of texts. So, it creates
the categories through an analysis of the degree of similarity of
the text documents that are to be classified. The similarity
measure is used to match these documents with pre-defined
categories [4] - [15]. Therefore, the document feature matrix is
formed to check that a document satisfies how many defined
features of the reduced feature set and categorized into which
category or class [4] [6] [7]. The fuzzy similarity measure can
be used to compute such different matrices.
III. THE FUZZY SIMILARITY BASED CONCEPT MINING
MODEL (FSCMM)
In this section, the proposed Fuzzy Similarity Based
Concept Mining Model (FSCMM) is discussed. This model
automatically classifies a set of known text documents into a
set of category groups. The model shows that how these
documents are trained step by step and classified by the
Support Vector Machine Classifier (SVMC). SVMC is further
used to classify various new and unknown text documents
categorically.
The proposed model is divided into the two phases: Text
Learning Phase (TLP) and Text Classification Phase (TCP).
TLP performs the learning function on a given set of text
documents. It performs the steps of first stage; i. e., Text
Document Training Processor (TDTP) and then the steps of
second stage; i. e., Fuzzy Feature Category Similarity Analyzer
(FFCSA). The TDTP is used to process the text document by
converting it into its small and constituent parts or chunks by
using NLP concepts at the Sentence, Document and Integrated
Corpora Levels. Then, it searches and stores the desired,
important and non-redundant concepts by removing stop
words, invalid words and extra words. In the next step, it
performs word stemming and feature reduction. The result of
sentence level preparation is low dimensional Reduced Feature
Vector (RFV). Each RFV of a document is sent for document
Vol. 2, No. 11, 2011
117 | P a g e
level preparation, so that Integrated Reduced Feature Vector
(IRFV) is obtained. To obtain IRFV, all the RFVs are
integrated into one. Now, Reduced Feature Frequency
Calculator (RFFC) is used to calculate the total frequency of
each different word occurred in the document. Finally, all
redundant entries of each exclusive word are removed and all
the words with their associated frequencies are stored in
decreasing order of their frequencies. At the integrated corpora
level, the low dimension Integrated Corpora Feature Vector
(ICFV) is resulted.
In such a way, feature vectors at each level are made low
dimensional by processing and updating step by step. Such
functionality helps a lot to search the appropriate concepts with
reduced vector length to improve system performance and
accuracy.
FFCSA performs similarity measure based analysis for
feature pattern (TD ICFV) using the enriched fuzzy logic
base. The membership degree of each feature is associated with
it. Therefore, an analysis is performed between every feature
of a text document and class.
SVMC is used for the categorization of the text documents.
It uses the concept of hyper planes to identify the suitable
category. Furthermore, SVMC accuracy is checked by
providing some new and unknown text documents to be
classified into the respective Category Group (CG). This task is
performed in TCP.
The proposed Fuzzy Similarity Based Concept Mining
Model (FSCMM) is shown in Fig. 1. In the next sections, this
model and its components are discussed in detail.

Figure 1. The Fuzzy Similarity Based Concept Mining Model (FSCMM).
Vol. 2, No. 11, 2011
118 | P a g e
A. Text Learning Phase (TLP)
Consider a set of n text documents,
TD = {TD
1
, TD
2
, TD
3
,,TD
n
} (1)
Where TD
1
, TD
2
, TD
3
,,TD
n
are the individual and
independent text documents which are processed, so that they
can be categorized into the required category.
1) Text Document Training Processor (TDTP)
Text Document Training Processor (TDTP) prepares the
given text document set TD of n text documents by performing
many operations on the sentence, document, and integrated
corpora levels. Firstly, each text document TD
i
, 1 i n, is
processed at its sentence level. The result of such sentence
level pre-processing for all the sentences of TD
i
is integrated
into one, which is further processed and refined to make
available for the integrated corpora. Integrated corpora accept
and integrate all the refined text documents and perform more
processing. Its result is sent to FFCSA.
a) At Sentence Level
This section describes that how a sentence is pre-processed
and finds out the feature vector, each features frequency and
finally, the conversion of the Feature Vector (FV) into the
Reduced Feature Vector (RFV).
A text document TD
i
is composed of a set of sentences, so
consider
TD
i
= {s
i1
, s
i2
, s
i3
,,s
im
} (2)
Where i denotes the text document number and m denotes
the total number of sentences in TD
i
. Sentence Extractor (SE) is
used to extract the sentence s
i1
from TDi. Each sentence has its
well-defined and non-overlapping boundaries, which makes the
sentence extraction a simple task for SE.
When the sentence is extracted, a verb argument structure
is made for s
i1
. The syntax tree is drawn using the pre-defined
syntactic grammar to separate the verbs and the arguments of
the sentence. The sentence can be composed of the nouns,
proper nouns, prepositions, verbs, adverbs, adjectives, articles,
numerals, punctuation and determiners. So, with the
construction of the syntax tree, the stop word and other extra
terms are removed except the nouns, proper nouns and
numerals which are considered as the concepts. To remove the
invalid and extra words, the Pseudo Thesaurus is used. It also
helps in word stemming.
The next step is to make the Feature Vector (FV) of the
sentence s
ij
of text document TD
i
as
FV = {F
i11
, F
i12
, F
i13
, . . . ,F
i1r
} (3)
Where 1 i n, 1 j m, and r depicts the total number of
present features in the s
ij
. Feature Frequency Calculator (FFC)
calculates the frequency of each different feature occurred in
FV. Frequency represents the number of occurrences of a
feature in the sentence. So, each different feature is associated
with its frequency in the form of a Feature Frequency pair as
<F
ijk
, freq (F
ijk
)>, where i is the text document number, j is the
sentence number of the sentence, k is the feature number, freq
() is a function to calculate the frequency of a feature, and 1 k
r.
The next step is to convert the high dimensional FV into
low dimensional Reduced Feature Vector (RFV) to reduce the
storage and execution time complexities. So, a counter loop is
invoked to remove the duplicate or redundant entries of a
feature. Therefore, only one instance of each different feature
occurred is stored in RFV. It highly reduces the FV dimension
and increases the efficiency of the system with good
performance. The complete sentence level processing is shown
in the Fig. 2.

Figure 2. Text Document Preparation at Sentence Level.
Such sentence level preprocessing for the TDP is performed
for each sentence of each document and then progressing
toward the integrated corpora.
b) At Document Level
This step accepts the resultant RFV which is sent for
document level pre-processing. Firstly, a counter loop is
invoked to match the similar features in each of the RFV
j
with
every other RFV
q
of a TD
i
where 1 j, q m, j q and 1 i
n. Refined Feature Vector Calculator (RFVC) updates each
features frequency for those features which are present two or
more times in more than two sentences. These updates are done
by adding up their frequencies in terms of combined calculated
frequency of that feature only. In this way, it updates the count
of each different occurred feature with more than one
occurrence. The features that have occurred only once in the
document will not update their frequency.
Vol. 2, No. 11, 2011
119 | P a g e
The next step is that all RFVs of a TD
i
are integrated into
one as
IRFV = Integrat (RFV
1
, RFV
2
,,RFV
m
) (4)
Where Integrat () is a function to combine all RFVs.
Now each RFV
j
is compared with every other RFV
q
where 1
j, q m, and j q. In this way, each feature of RFV
j
is
compared with the every other feature of the RFV
q
and thereby,
the duplicate and redundant features are removed. The
complete procedure on the document level is shown in Fig. 3.

Figure 3. Text Document Preparation at Document Level.
c) At Integrated Corpora
In integrated corpora, all the IRFVs of n documents are
integrated into one. This step is used to calculate and update the
final frequency of each different feature occurred in the
corpora. Firstly, it removes duplicate and redundant feature
entries of IRFV, and then removes all ambiguous words with
the help of the pseudo thesaurus. In such a way, Integrated
Corpora Feature Vector (ICFV) is generated. A Threshold
Value (TV) is defined for ICFV. TV cuts off those features
whose total frequency is less than TV. Finally, it reduces the
dimension of ICFV.
Therefore, Integrated Corpora Feature Vector (ICFV) is
constructed as
ICFV = {F
1
, F
2
,,F
n
} (5)
Where F
1
, F
2
,,F
n
represent the features.
Such features with their associated frequencies show the
statistical similarity between the text documents. Each feature
is counted for each document and represented as the feature
and its frequency as follows.
TD
1
= {<F
1
, f (F
1
)>, <F
2
, f (F
2
)>,,<F
n
, f (F
n
)>} (6)

Where f(F
i
) is the function to calculate the frequency of
feature F
i
in the text document.
The next step is to generate the matrix of TD and ICFV in
the given form of table 1. In this matrix, the 0 represents the
absence of that feature in TD and the numerical value
represents the total number of occurrences of the feature in the
TD.
TABLE I. TD ICFV MATRIX

Text
Document
Feature
F1 F2 F3
TD1 1 0 1
TD2 1 3 0
TD3 0 2 1
TD4 4 0 0

2) Pseudo Thesaurus
The Pseudo Thesaurus is a predefined English
Vocabulary Set which is used to check the invalid words or to
remove extra words from a sentence while processing the
sentence in the TDTP. It is also used for word stemming so that
the exact word can be obtained. For example, consider three
different words for the word research - researching, researcher
and researches. When the word stemming is performed,
research is the final resulting word with the feature frequency
counted as 3.
3) Class Feeder (CF)
Text Classification is the process of assigning the name of
the class to a particular input, to which it belongs. The classes,
from which the classification procedure can choose, can be
described in many ways. So classification is considered as the
essential part of many problems solving tasks or recognition
tasks. Before classification can be done, the desired number of
classes must be defined.
4) Fuzzy Fetaure Category Similarity Analyzer (FFCSA)
In FFCSA, firstly the Feature Pattern (FP) is made in the
form of the membership degree of each feature with respect to
every class. Consider text document set TD of n text documents
as per given in equation 1, together with the ICFV F of y
features f
1
, f
2
,f
y
and e classes c
1
, c
2
, , c
e
. To construct the FP
for each feature f
k
in F, its FP fp
i
is defined, by
fp
i
= < fp
i1
, fp
i2
, fp
i3
,, fp
ie
> (7)
= < (f
i
, c
1
), (f
i
, c
2
), (f
i
, c
3
),, (f
i
, c
e
)>
Where,
f
i
represents the number of occurrences of f
i
in the text
document TD
g
where 1 g n.
(f
i
, c
e
) is defined as the sum of product of the feature
value of f
i
present in n text documents TD, w. r. t. a column
vector and the 1 or 0 as the presence or absence of that feature
in class c
e
/ Sum of the feature value for the class c
e
only,
where 1 j e. So,
(f
i
, c
e
)
()
()
(8)

b
i
is represented as
b
i
= 1, if document class c
e
(9)
= 0, otherwise
Each text document TD belongs to only one class c. In this
way, each class can belong to one or more text documents. A
Vol. 2, No. 11, 2011
120 | P a g e
set of n documents and their related categories or classes are
represented as an ordered pair as shown
TD = {<TD
1
, C (TD
1
)>, <TD
2
, C (TD
2
)>,,<TD
m
, C
(TD
m
)>} (10)
Where the class of the text document TD
i
: C(TD
i
) C, C
(TD) is a categorization function whose domain is TD and
range is C.
Each document belongs to one of the classes in the C
(TD).The resulted text documents that have many features are
stored with their relevant classes as shown in the table 2. They
are in the form of <doc no, number of occurrences of each
feature, class no>.
TABLE II. TEXT DOCUMENT FEATURE VECTOR CLASS MATRIX

Text
Document
Feature
Class
F1 F2 F3

TD1 1 0 1
C1
TD2 1 3 0
C2
TD3 0 2 1
C2
TD4 4 0 0
C3

In such a way, the relation between a feature and a class is
made. Sometimes, it is quite possible that one document
belongs to two or more classes that concern has to be
considered by making the more presences of the text document
in the table with the cost of increased complexity, so it is
required to check each features distribution among them.
B. Support Vector Machine Classifier (SVMC)
The next step is to use the Support Vector Machine
Classifier (SVMC). SVMC is a popular and better method than
other methods for text categorization. It is a kernel method
which finds the maximum margin hyper plane in the feature
space paradigm separating the data of training patterns into two
groups like Boolean Logic 1 and 0. If any training pattern is not
correctly classified by the hyper plane, then the concept of
slack measure is used to get rid out of it.
Using this idea, SVMC can only separate apart two classes
for h = +1 and h = -1. For e classes, one SVM for each class is
created. For the SVM of class c
l
, 1 l e, the training patterns
of class c
l
are for the h = +1 and of other classes are h = -1.
The SVMC is then the aggregation of these SVMs.
SVM provides good results then KNN method because it
directly divide the training data according to the hyper planes.
C. Text Classification Phase
To check the predictive accuracy of the SVMC, new and
unknown text document is used, which is independent of the
training text documents and is not used to construct the SVMC.
The accuracy of this document is compared with the learned
SVMCs class. If the accuracy of the SVMC is acceptable and
good, then it can be used further to classify the future unseen
text documents for which the class label is not known.
Therefore, they can be categorized into the appropriate and a
suitable category group.
IV. CONCLUSION AND FUTURE SCOPE
Text classification is expected to play an important role in
future search services or in the text categorization. It is an
essential matter to focus on the main subjects and significant
content. It is becoming important to have the computational
methods that automatically classify available text documents to
obtain the categorized information or groups with greater speed
and fidelity for the content matter of the texts.
As the proposed FSCMM model is made for text document
categorization, it works well with high efficiency and
effectiveness. Although this model and methodology seem very
complex, yet it achieves the task of text categorization with
high performance, and good accuracy and prediction. Feature
Reduction is performed on the sentence, document and
integrated corpora levels to highly reduce feature vector
dimension. Such reduction improves the system performance
greatly in terms of space and time. Result shows that the
feature reduction reduces the space complexity by20%.
Fuzzy similarity measure and methodology are used to
make the matching connections among text documents, feature
vectors and pre-defined classes. It provides the mathematical
framework for finding out the membership degrees as feature
frequency.
This model shows better results than other text
categorization techniques. SVM classifier gives better results
than KNN method. The system performance does not show
high information gain and prediction results when KNN is used
because it produces noise sensitive contents.
In the future, such model can be further extended to
include the non-segmented text documents. It can also be
extended to categorize the images, audio and video - related
data.
REFERENCES
[1] N. P. Padhy, Artificial Intelligence and Intelligent Systems, 5
th
ed.,
Oxford University Press, 2009.
[2] Eliane Rich, Kevin Knight and Shivashankar B Nair, Artificial
Intelligence, 3
rd
ed., Mc Graw Hill, 2010.
[3] Jiawei Han, and MicheLine Kamber, Data Mining: Concepts and
Techniques, 2
nd
ed., Elsevier, 2006.
[4] Marcus Vinicius, C. Guelpeli, Ana Cristina, and Bicharra Garcia, An
Analysis of Constructed Categories for Textual Classification Using
Fuzzy Similarity and Agglomerative Hierarchical Methods, Third
International IEEE Conference Signal-Image Technologies and Internet-
Based System, September 2008.
[5] S. Shehata, F. Karray, and M. S. Kamel, An Efficient Concept-Based
Mining Model for Enhancing Text Clustering, IEEE Transactions On
Knowledge And Data Engineering, Vol. 22, No. 10, October 2010.
[6] Jung-Yi Jiang, Ren-Jia Liou, and Shie-Jue Lee, A Fuzzy Self-
Constructing Feature Clustering Algorithm for Text Classification,
IEEE Transactions On Knowledge And Data Engineering, Vol. 23, No.
3, March 2011.
[7] Choochart Haruechaiyasak, Mei-Ling Shyu, Shu-Ching Chen, and Xiuqi
Li, Web Document Classification Based on Fuzzy Association, IEEE,
2007.
[8] Ahmad T. Al-Taani, and Noor Aldeen K. Al-Awad, A Comparative
Study of Web-pages Classification Methods using Fuzzy Operators
Vol. 2, No. 11, 2011
121 | P a g e
Applied to Arabic Web-pages, World Academy of Science,
Engineering and Technology, 2005.
[9] Qing YANG, Wei CHEN, and Bin WEN, Fuzzy Ontology Generation
Model using Fuzzy Clustering for Learning Evaluation, 2008.
[10] Kok WaiWong, Todsanai Chumwatana, and Domonkos Tikk,
Exploring The Use of Fuzzy Signature for Text Mining, IEEE, 2010.
[11] O. Dehzangi, M. J. Zolghadri, S. Taheri and S.M. Fakhrahmad,
Efficient Fuzzy Rule Generation: A New Approach Using Data Mining
Principles and Rule Weighting, Fourth International Conference on
Fuzzy Systems and Knowledge Discovery (FSKD), 2007.
[12] Surya Sumpeno, Mochamad Hariadi, and Mauridhi Hery Purnomo,
Facial Emotional Expressions of Life-like Character Based on Text
Classifier and Fuzzy Logic, IAENG International Journal of Computer
Science, May, 2011.
[13] Rahmah Mokhtar, Siti Norul Huda Sheikh Abdullah, and Nor Azan Mat
Zin, Classifying Modality Learning Styles based on Production Fuzzy
Rules, International Conference on Pattern Analysis and Intelligent
Robotics, June 2011.
[14] Giuseppe Fenza, Vincenzo Loia, and Sabrina Senatore, Concept
Mining of Semantic Web Services By Means Of Extended Fuzzy Formal
Concept Analysis (FFCA), IEEE, Feb. 2008.
[15] Chen Yanyun , Qiu Jianlin, Gu Xiang,Chen Jianping, Ji Dan, and Chen
Li, Advances in Research of Fuzzy C-means Clustering Algorithm,
International Conference on Network Computing and Information
Security, 2011.
AUTHORS PROFILE
Shalini Puri is pursuing M.Tech in Computer Science at
Birla Institute of Technology, Mesra, Ranchi, India. She
is currently working as an Assistant Professor in a reputed
engineering college in India. Her research areas include
Artificial Intelligence, Data Mining, Soft Computing,
Graph Theory, and Software Engineering.

Vol. 2, No. 11, 2011
122 | P a g e
Improved Echo cancellation in VOIP

Patrashiya Magdolina Halder
Dept. of Electronics and Telecommunication Engineering
Daffodil International University
Dhaka, Bangladesh
A.K.M. Fazlul Haque
Dept. of Electronics and Telecommunication Engineering
Daffodil International University
Dhaka, Bangladesh

Abstract VoIP (voice over internet protocol) is very popular
communication technology of this century and has played
tremendous role in communication system. It is preferred by all
because it deploys many benefits it uses Internet protocol (IP)
networks to deliver multimedia information such as speech over a
data network. VoIP system can be configured in these connection
modes respectively; PC to PC, Telephony to Telephony and PC to
Telephony. Echo is very annoying problem which occurs in VoIP
and echo reduces the voice quality of VoIP. It is not possible to
remove echo 100% from echoed signal because if echo is tried to
be eliminated completely then the attempt may distort the main
signal. That is why echo cannot be eliminated echo perfectly but
the echo to a tolerable range. Clipping is not a good solution to
suppress echo because part of speech may erroneously removed.
Besides an NLP does not respond rapidly enough and also
confuses the fading of the voice level at the end of a sentence with
a residual echo. This paper has proposed echo cancellation in
VoIP that has been tested and verified by MATLAB. The goal
was to suppress echo without clipping and distorting the main
signal. By the help of MATLAB program the echo is minimized
to enduring level so that the received signal seems echo free. The
percentage of suppressing echo varies with the amplitude of the
main signal. With regarding the amplitude variation in received
(echo free) signal the proposed method performs better in finding
the echo free signal than the other conventional system.
Keywords- PSTN; round trip delay; impedance; inverse filtering;
denoise; histogram amplifier; repeater.
I. INTRODUCTION
Around 20 years of research on VoIP, some problems of
VoIP are still remaining and a substantial problem in
telecommunications is the generation of echo. Echo is a
phenomenon where a delayed and distorted version of an
original signal is reflected back to the source. Echo is a
congenital problem which mainly occurs in PSTN (Public
Switching Telephone Network) [1]. Echo occurs in analogy
part of a telecommunication system .Echo is generated by
human voice is heard as they are reflected from the floor,
walls and other neighboring objects. If a reflected wave arrives
after a very short time of direct sound, it is considered as a
spectral distortion or reverberation. However, when the
leading edge of the reflected wave appears again a few tens of
milliseconds after the direct sound then it is heard as an
audible echo [2]. Echo is annoying when the round trip delay
exceeds 30 ms. such an echo is typically heard as a hollow
sound. Echoes must be loud enough to be heard. Echo which
is less than thirty 30 dB is rarely to be noticed. But when
round trip delay exceeds 30 ms and echo strength exceeds 30
dB, echoes become steadily more disruptive. Every echo does
not reduce voice quality. There are mainly two kinds of echo,
that is Hybrid echo and Acoustic echo. Hybrid echo, line echo
or electrical echo is different names given to the echo
generated by an impedance mismatch in the analog local loop.
The impedance mismatch occurs when the two-wire network
meets the four-wire network [3]. The impedance of subscriber
lines vary from one subscriber to the next, this time sharing
makes it impossible to provide a perfect impedance match for
every line [4]. Acoustic echo is caused by acoustic coupling
problems between a telephone's speaker and its microphone.
Acoustic echo can occur in mobile phones, wire line
telephones or in a hands-free set of a speaker phone [4]. It can
be caused by hand-set crosstalk in poor quality handsets or by
echo in the environment surrounding the caller. There are
some works on frequency reuse scheme [3, 6-9].
According to asterisk echo cancellation previously called
carbon profile [6, 7] is operated by generating multiple copies
of the received signal, each delayed by some small time
increment. These delayed copies are then scaled and
subtracted from the original received signal.Srinivasaprasath
Raghavendran et al [3] has proposed an echo cancellation
process using MATLAB but there the far end signal and the
near end signal is taken separately and then tested whether
there is echo or not by Double talk detector. This process also
includes NLMS and subtraction .Ganesan Periakarruppan,
Andy L.Y.Low, Hairul Azhar b Abdul Rashid et al [8]
introduced that PBEC the sample to generate the echo replica
model will be used to subtract the. Jerker Taudien el al [9]
suggested Line probing is a method of inserting a known
signal at the far-end and recording the near-end signal. The
two signals are then analyzed together for various
impediments. Three tone sweeps of different power levels are
used to probe the line in the non-linear distortion analysis tool.
The tone sweeps are recorded in three different power levels to
detect clipping.
All these procedures require more things than this
proposed method. This paper has suggested cancelling echo
from echoed signal. Only received signal is analysed here so it
is not needed to analyse near end and far end signal separately.
Besides subtraction and clipping is not required here which ma
affect the main signal. This is a very simple program which
eliminates echo. Inverse filtering is used here which analyse
the received signal and remove echo from the acquired signal.

Vol. 2, No. 11, 2011
123 | P a g e
0 2 4 6 8 10 12 14 16 18
x 10
4
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
0 0.5 1 1.5 2 2.5 3
x 10
5
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18
x 10
4
-1.5
-1
-0.5
0
0.5
1
1.5
II. BACKGROUND
Within the caller's telephone, a certain amount of the signal
from the microphone is fed straight back to the earpiece. An
improperly balanced hybrid won't correctly filter out the entire
transmitted signal, and will reflect some of it back down the
other half of the trunk. Imbalance may be from poor design
(common) or unpredictable. The reasons of echo are as
follows:
1. Poor room acoustics
2. Marginal microphones for soft terminals
3. Low quality cellular handsets
4. Deficient echo control in the terminal device its
5. Bridge-taps (something done by the Telco, seldom seen any
more)
6. Use of lengthy untwisted wire within the subscriber's
premises
III. SIMULATION AND RESULTS
This program is to explore the problem of echo
cancellation. Iinverse filtering (IF) is a widely known method
for voice and speech analysis, which mainly works on
estimating the source of voiced speech. This method enables
to estimate the glottal volume velocity waveform or glottal
airflow. The idea behind inverse filtering is to form a
computational model for glottal pulse detection by filtering the
speech signal.

Figure 1. Flowchart for echo cancellation using MATLAB
This flowchart is given bellow for echo suppression using
MATLAB (Fig. 1) .The main signal, signal with echo and
signal without echo is showed respectively in Fig. 2, 3 and 4.
This signal can be denoised using Matlab to remove echo from
this signal.
For the experiment at first voice signal is acquired, it can
be done by any kind of speech recorder. Then this acquired
voice signal is used to create a .wave file using the audio
signal. After that Matlab software is opened the data of the
signal is imported to create .mat file. This mat file has to be
loaded to transfer the value of the speech signal in workspace.
To remove echo from the signal wavelet is used. To inspect
the signals by wavelet wave menu is typed in the command
window of Matlab; a new window of wavelet will be
displayed then. The echoed signal is loaded in wavelet and
then the signal is analyze, view the statistics, denoise and
analyze the signal. Doing all these signal with echo can be
analyzed. The signals can be compressed to get better result

Figure 2. Main signal

Figure 3. Signal with echo

Figure 4. Signal without echo
By running the program, hearing the signal it will be clear
that echo is removed from the echoed signal and the quality of
signal is improved. The signal is clear and each part of speech
is present. If the simulation figures are observed closely then it
can be seen that the wave of main signal is like overlapped in
the echoed signals figure. The refined signal is almost look
alike the echo free signal, there is just change in the amplitude
of the signal. So it can be said that the improved signal and
the outcome signal both prove that the attempt is successful to
remove echo and perform better.
Histogram is used to analyze the speech signal which
showed in Fig. 5, 6 and 7.

Vol. 2, No. 11, 2011
124 | P a g e
Figure 5. Analyzing main signal with wavelet
Figure 6. Analyzing signal with echo with wavelet
Figure 7. Analyzing signal without echo with wavelet
The following figure 8 shows the difference among the
signals with the help of FFT (Fast Fourier Transform)
Figure 8. Analyzing all signals using FFT
Difference between signal with echo and signal without
echo is observed. By seeing the waveforms generally it can be
said that echo is removed. There is difference in amplitude
among the signals; the amplitude can be regained by using
amplifier or repeater.
For further analyzing different statistical values are
extracted from this experiment and taking the values a table
4.4 and a graph (Fig. 9) is plotted, this will help to prove that
echo is removes from the echoed signal.
TABLE I. SIGNAL ANALYZING

Main signal Echoed signal

Signal without
echo
Mean -7.236
e-008
0.7844

4.797e-
007
0.999 1.432e-
008
1.13
Median -
0.0004
578
-0.8289

-6.104e-
005
-
0.841
2
-
0.000589
-
1.10
6
Mode 0.0046
48
1.613 -0.01373 1.839 -0.02528 2.23
7
Standard
deviation
0.1336

0.1186 0.1918
Median
absolute
deviation
0.01762

0.01541 0.05687
Mean
absolute
deviation
0.07034

0.0599 0.1211
Vol. 2, No. 11, 2011
125 | P a g e
Figure 9: Comparing main signal, echoed signal and signal without echo.
Difference between the signal without echo and signal with
echo is observed. The mean, median, mode all the values are
analyzed more over the compressed signal is almost matched
with the desire echo free signal , which is our goal.
IV. CONCLUSION
In this report, empirical audio signal has been considered
to evaluate the performance of echo cancellation. It has been
observed that the echo is suppressed without changing or
distorting the main signal and user of VOIP can hear clear
sounds. This program is easy to use and simpler than the
conventional methods of suppressing echo. The simulation
results have been tested and verified using Wavelet Tool.
Power spectrum density has also been used to observe the
difference of the received signal. The proposed system has
found better performance in finding the echo cancellation than
the conventional methods which can be adopted to suppress
echo and take the fullest advantage of VOIP telephony.
REFERENCES
[1] A Survey on Voice over IP over Wireless LANs Haniyeh Kazemitabar,
Sameha Ahmed, Kashif Nisar, Abas B Said, Halabi B Hasbullah
[2] Echo in Voice over IP Systems Series VoIP Performance Management
Date January 2006
[3] Implementation of an Acoustic Echo Canceller Using Matlab by
Srinivasaprasath Raghavendran College of Engineering University of
South Florida October 15, 2003
[4] IP Telephony: Packet-based multimedia communication system By
Olivier Hersent, David Gurle, Jean-Pierre Petit
[5] Design a high-performance echo canceller for VOIP application, by Dr.
Chang Y. Chob
[6] http://www.voipinfo.org/wiki/view/Asterisk+echo+cancellation
[7] http://www.eetimes.com/design/signal-processing-dsp/4017606/High-
Performance-echo-canceller-for-Asterisk-VoIP-systems
[8] PACKET BASED ECHO CANCELLATION FOR VOICE OVER
INTERNET PROTOCOL SIMULATED WITH VARIABLE AMOUNT
OF NETWORK DELAY TIME by Ganesan Periakarruppan, Andy
L.Y.Low, Hairul Azhar b Abdul Rashid
[9] LINE PROBING IN VOIP NETWORKS TO FIND PERFORMANCE
LIMIT OF ECHO CANCELLER by Jerker Taudien, 2007

-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
Mean Median Mode Standard
deviation
Median
absolute
deviation
Mean
absolute
deviation
main
echoed
echo free
Vol. 2, No. 11, 2011
126 | P a g e
A New Test Method on the Convergence and
Divergence for Infinite Integral

Guocheng Li
Linyi University at Feixian
Feixian, Shandong, P.R.China

AbstractThe way to distinguish convergence or divergence of
an infinite integral on non-negative continuous function is the
important and difficult question in the mathematical teaching all
the time. Using the comparison of integrands to judge some
exceptional infinite integrals are hard or even useless. In this
paper, we establish the exponential integrating factor of negative
function, and present a new method to test based on its
exponential integrating factor. These conclusions are convenient
and valid additions of the previously known results.
Keywords-Iinfinite integral; Exponential integrating factor;
Convergence and divergence.
I. INTRODUCTION
The convergence and divergence of infinite integral plays a
significant role in mathematical analysis, and has been
received much attention of many researchers. Many effective
methods have been proposed such as Ordinary Comparison
Test, Limiting Test, Dirichlet Test and Abel Test [1]. The basic
ideas of these methods are to find another comparison infinite
integral that its convergence or divergence is certain. But it is
difficult or impossible to find the comparison infinite. So these
arouse great interest of many scholars to research ([2-3]).
It is well-known that the convergence and divergence of
infinite integral for the different integrand which its limit is
zero is different. Base on the geometric meaning of infinite
integral, it equal to the size of the area which is surrounded by
the integrand (image) and X axis (not closed). The
Convergence and divergence mainly determined by the
proximity degree of integrand tends to the x-axis. In this paper,
we discuss the test method for the convergence and divergence
of infinite integral
( )
a
f x dx
+
} ( ) f x
is the
nonnegative continuous function which is defined in the
interval
[ , ), 0 a a + >
the problem by
ln ( ) / ln f x x
[3], it can't intuitive
describe the proximity degree of integrand tends to the x-axis,
so the exponential integrating factor of negative function is
created and we obtain a new test method for the convergence
and divergence of infinite integral. This method is more simple
and feasible than ever before.
II. CHARACTERIZATIONS ON CONVERGENCE OF INFINITE
INTEGRAL
Definition 2.1 Assume that ( ) f x is non-negative
differentiable function, we call the formula
'
( ) ( ) / ( ) R x f x f x = is the function of the exponential
integrating factor of ( ) f x , abbreviated as factor function.
From definition 2.1, we suppose the function

0
0 0
( ) ( ) exp ( ) ,
x
x
f x f x R t dt x a

= >
`

)
}
is completely
determined by its factor function ( ) R x . When the
independent variable tend to be infinite, we can use the size of
factor function to compare the different functions reduce speed
of which tend to the x-axis. For lim ( ) 0
x
f x
+
= is the
necessary condition of convergence for infinite integrals
( )
a
f x dx
+
}
([1]), so we can easy to push that ( ) 0 R x > is
the necessary conditions of convergence for ( )
a
f x dx
+
}
.
If
( )
lim 0
( )
x
f x
c
g x
+
= > , we have ( ) f x and ( ) g x are
equivalent. We denoted by ~
t
f g .The nature of the
equivalent functions is incredibly similar, so what about the
infinite integration convergence of the equivalent integrand?
We have the following conclusion.
Theorem 2.1 Assume that Non-negative differentiable
functions ( ) f x and ( ) g x are equivalent, then its
convergence and divergence is incredibly similar.
Proof: By
( )
( )
lim lim 0
( )
( )
x
a
x
x x
a
f t dt
f x
c
g x
g t dt

= = >
}
}
,

Vol. 2, No. 11, 2011
127 | P a g e
There exist constant , 0 m M > such that
( ) / ( ) , m f x g x M s s . x a > Obviously, we have
( ) ( ) ( ) ;
1 1
( ) ( ) ( )
a a a
a a a
m g x dx f x dx M g x dx
f x dx g x dx f x dx
M m
+ + +
+ + +
s s
s s
} } }
} } }

From the definition of convergence of the infinite integrals,
the desired result follows.
Theorem 2.2 Assume that ( ) f x , ( ) g x are non-
negative differentiable functions in interval [ , ) a + and
( ) / ( ) f x g x is finally monotone decreasing, then
(i)
0 0
, , ( ) ( )
f g
x x x R x R x - > > .
(ii) If ( )
a
g x dx
+
}
is convergent, then ( )
a
f x dx
+
}
is
convergent, If ( )
a
g x dx
+
}
is divergent, then ( )
a
f x dx
+
}
is
divergent.
Proof:
(i)
0 0
, x x x - >
' '
'
2
( ) ( ) ( ) ( ) ( )
( )
( ) ( )
1
[ ( ) ( )] 0
( ) ( )
( ) ( ).
g f
f g
f x g x f x f x g x
g x g x
R x R x
f x g x
R x R x
= =
<
>

(ii)
( )
0, ,
( )
f x
M let M
g x
- > s then
( ) ( )
a a
f x dx M g x dx
+ +
s
} }
. Using Ordinary Comparison
Test, we obtain this conclusion.
Assume that we know the convergence and divergence of
an infinite integral, we can obtain the convergence and
divergence of other infinite integrals by comparing the size of
the exponential factor. But it's complex. In fact, it can be
derived alone by the exponential factor ( ) R x of integrand.
Theorem 2.3 Assume that ( ) f x is non-negative
differentiable function in [ , ) a + ,
(i) If lim ( ) 1,
x
xR x
+
> then ( )
a
f x dx
+
}
is convergent,
(ii) If lim ( ) 1,
x
xR x
+
< then ( )
a
f x dx
+
}
is divergent.
Proof: (i) 0, k > we have
'
( ) 1
,
( )
f x k
f x x
+
> then
1
ln ( ) ln( ),
k
f x cx
+
> 0 c > .
That is
(1 )
1
( )
k
f x x
c
+
s i.e., According to the Ordinary
Comparison Test, ( )
a
f x dx
+
}
is convergent.
(ii) Let
1
( ) , g x
x
= then
1
a
dx
x
}
is divergent, for
1
( ) ( ),
g
R x R x
x
< = from Theorem 2.2, we have that
( )
a
f x dx
+
}
is divergent.
Combining [3], we have that ( )
a
f x dx
+
}
is possible
convergent or divergent if lim ( ) 1
x
xR x
+
= from some
possible examples. In fact, we can continue to study
convergence as follow:
Theorem 2.4 Assume that ( ) f x is non-negative
differentiable function in interval[ , ) a + and
lim ( ) 1
x
xR x
+
= .
(i) If lim ln [ ( ) 1] 1
x
x xR x
+
> , then ( )
a
f x dx
+
}
is
convergent,
(ii) If lim ln [ ( ) 1] 1
x
x xR x
+
s , then ( )
a
f x dx
+
}
is
divergent.
Proof: (i) We can obtain that ( ) R x is monotone-non-
increasing function, for all 0 k > , we have
{ }
0
0
0
0 1
( ) exp ( )
1
exp (1 )
ln
1
exp
ln
k
f x c xR x
k
c
x
c k
c
x x
+
s s
+
+ <
`
)
+
=
`
)

From Ordinary Comparison Test we can obtain that
( )
a
f x dx
+
< +
}

Vol. 2, No. 11, 2011
128 | P a g e
(ii)Method1: Let
1 ln
( ) ,
ln
x
R x
x x
+
s we have that
ln ( ) ln( ln ) f x x x s , i.e.
1
( )
ln
f x
x x
> , by
1
ln
a
dx
x x
}

is divergent, combining Ordinary Comparison test, we can
obtain that ( )
a
f x dx
+
}
is divergent.
Method2: Using the condition, we obtain that
1 ln
( ) ( )
ln
g
x
R x R x
x x
+
s = ,
Where
1 1
( ) ,
ln ln
a
g x and dx
x x x x
=
}
is divergent.
From theorem2.2, we have that ( )
a
f x dx
+
}
is divergent too.
Using the conclusions above, we can easily think of the
following examples.
Example 1: Discuss the convergence of
2
1
ln
p q
dx
x x
}
.
Solution: For its exponential integrating factor
( )
c
p cqx
R x
x
+
= ,
We easy to know that
ln
lim
ln
x
p x q
x p
x x
+
+
= , using
theorem 2.3, we have that the integration is convergent if
1 p > , and is divergent if 1 p < ; when 1 p = , we easy to
know [ ( ) 1] x R x q = , using theorem 2.4, we get that the
integration is convergent when 1 q > ; the integration is
divergent when 1 q s .
Example2: Discuss the convergence of
1
exp( )
p c
x qx dx
}
.
Solution: For its exponential integrating factor is
( )
c
p cqx
R x
x
+
= ,
We easy to know that lim ( ) [ ]
c
x
xR x p cqx
+
= + , using
theorm2.3, we have
1
exp( )
p c
x qx dx
}
is convergence if
lim 1
c
x
p cqx
+
+ < , i.e. 0 c s or 0, 1 q p = < or
c>0, q<0
.
III. CONSTRUCTING CONVERGENT INFINITE INTEGRAL AND
POSSIBLE APPROACH
From Theorem 2.4, we can obtain that the convergence and
divergence of ( )
a
f x dx
+
}
is determined by the exponential
integrating factor of ( ) f x .
Suppose that ( )
a
f x dx
+
}
and ( ) ( )
a
f x g x dx
+
}
are
convergent, we give the following conditions of the non-
negative function ( ) g x .
Using the exponential integrating factor of ( ) ( ) f x g x ,
i.e. ( ) ( ) ( )
fg f g
R x R x R x = + ,
if lim( [ ( ) ( )]) 1
f g
x
x R x R x
+
+ > , then
( ) ( )
a
f x g x dx
+
}
is convergent. In fact, we can construct
some convergent infinite integral for some specific needs.
Theorem 3.1 Assume that ( )
a
f x dx
+
}
is convergent, if
the non-negative function ( ) g x is monotonous and bounded,
then ( ) ( )
a
f x g x dx
+
}
is convergent.
Proof: Combining the conditions and theorem 2.4, we
know that lim ln [ ( ) 1] 1
x
x xR x
+
> . If ( ) g x is decreasing
and lower bounded,
Then ( ) 0
g
R x > and ( ) ( ) ( )
fg f g
R x R x R x = + hold.
Obviously, lim( [ ( ) 1]) 1
fg
x
x R x
+
> ; If ( ) g x is increasing
and upper bounded, then ( )
x
g
a
R t dt
}
is convergent, i.e.
lim ln ( ) 0
g
x
x xR x
+
= , such that
lim ln [ ( ) 1] lim ln [ ( ) 1] 1
fg f
x x
x xR x x xR x
+ +
= > .
From all above, we can obtain that ( ) ( )
a
f x g x dx
+
}
is
convergent.
Theorem 3.2 Suppose ( )
a
f x dx
+
}
is convergent where
( ) f x is continuous and nonnegative in [ , ) a + .
(i)If < and 1, o < then ( )
a
x f x dx
o
}
is convergent.
(ii)If , , R o = e then ( )
a
x f x dx
o
}
is convergent,
Vol. 2, No. 11, 2011
129 | P a g e
(iii)If = and
1
1
ln[ [ ( ) ] 1]
lim ,
ln
x
x
R t dt
t
x
|
+
+
<
}

Then
{ }
exp ( )
a
x f x dx
|
}
is convergent, where
lim inf ( )
x
xR x
+
= ,
Proof: (i) For ( ) R x
x
o
is the factor function of
( ) x f x
o
, if 1 o < , then lim ( ) 1
x
xR x o
+
> ,i.e.
lim [ ( ) ] 1
x
x R x
x
o
+
> ,
Using theorem 2.3, we get ( )
a
x f x dx
o
}
is convergent.
(ii) Easy to get by (i).
(iii) From condition
1
1
[ ( ) ] 1
x
x R t dt
t
|
< +
}
, i.e.
1
1 1
1
[ ( ) ]
x x
t dt R t dt
t
|
|

<
} }
,
We have
1
1
( ) x R x
x
|
|

< , from
{ }
exp ( ) x f x
|
, we
obtain its factor function
1
( ) R x x
|
|

, using theorm2.3, we
get
1
lim [ ( ) ] 1
x
x R x x
|
|

+
> . Base on L'Hospital rule, we
have
1
( ) 1
lim
1
[ ( ) ] 1
x x
xR x
R t dt
t
|
+
<
+
}
.
Thus, we have that
{ }
exp ( )
a
x f x dx
|
}
is convergent.
IV. CONCLUSIONS AND PROSPECT
In this paper, we get a new test method for the convergence
and divergence of infinite integral, as everyone knows, defect
integral can be changed into infinite integral by
1
x
t
=
so we can
expands our method on the defect integral.
We can almost judge the convergence of the infinite
integrals by the conclusions refer to in Section 2. So we
discuss some questions of infinite integral by the properties of
the convergent infinite integral. In fact, the most important
application in Section 3 is to construct plenty of distributions
in the probability field especially in the Risk Theory. The
claim distribution is vital for both the insured and the insurance
company [4-7], constructing suitable distributions is beneficial
to need for the insured or development of the insurance
companies. This is a topic for future research.
ACKNOWLEDGMENT
The authors wish to give their sincere thanks to the
anonymous referees for their valuable suggestions and helpful
comments which improved the presentation of the paper.
REFERENCES
[1] East China normal university, Mathematical analysis [M], Higher
education press, 2001. (In Chinese)
[2] Wen Z.Y., Convergence of the infinite integrals of a new identification
method [J], Mathematics,2005.21(2):111-112. (In Chinese)
[3] Luo L.P., The discussion of the necessary conditions for infinite integral
convergence [J], Advanced mathematics research, 2005,8(4):19-21. (In
Chinese)
[4] Embrechts P, Veraverbeke N. Estimates for the propabilty of ruin with
special emphasis on the possibility of large claims. Insurancee Math
Econ, 1982, 1: 55-72
[5] Klppelberg C., Subexponential distributions and characterizations of
relateed class. Probab. Th. Rel.Fields,1989, 82: 259-269.
[6] Cline D B H. Convolutions of distributions with exponential and
subexponential tails. Austral Math Soc (Series A), 1987, 43: 347-365
[7] Yin Chuancun. A local theorem for the probability of ruin. Science in
China (Series A). 2004, 47: 711-721.

Vol. 2, No. 11, 2011
130 | P a g e
Adaptive Outlier-tolerant Exponential Smoothing
Prediction Algorithms with Applications to Predict
the Temperature in Spacecraft

Hu Shaolin
State Key Laboratory of Astronautics
Xian,P.O.Box 505-16,710043,China
Zhang Wei
State Key Laboratory of Astronautics
Xian,P.O.Box 505-16,710043,China
Li Ye
School of Automation,
Xian University of Technology, 710048, China
Fan Shunxi
School of Automation,
Xian University of Technology, 710048, China

AbstractThe exponential smoothing prediction algorithm is
widely used in spaceflight control and in process monitoring as
well as in economical prediction. There are two key conundrums
which are open: one is about the selective rule of the parameter
in the exponential smoothing prediction, and the other is how to
improve the bad influence of outliers on prediction. In this paper
a new practical outlier-tolerant algorithm is built to select
adaptively proper parameter, and the exponential smoothing
prediction algorithm is modified to prevent any bad influence
from outliers in sampling data. These two new algorithms are
valid and effective to overcome the two open conundrums stated
above. Simulation and practical results of sampling data from
temperature sensors in a spacecraft show that this new adaptive
outlier-tolerant exponential smoothing prediction algorithm has
the power to eliminate bad infection of outliers on prediction of
process state in future.
Keywords-Exponential prediction; Adaptive smoothing prediction;
Outlier-tolerance smoothing prediction.
I. INTRODUCTION
The exponential smoothing prediction algorithm was first
suggested by Brown as an operation research analyst during
World War II, when he worked on submarine locating and
tracking model for antisubmarine warfare. He thought the trend
of time series was stability or regularity, so the future state of
time series could be deduced reasonably [1]. After more than
half a century, exponential smoothing forecasting method has
become a common forecasting method as well as the most
commonly used forecasting method in short-term economic
development trend forecast and many other areas in [1,2].
Exponential smoothing prediction method has obvious
advantages: one is that the algorithm is simple, the predicted
value at time
1 + k
t
can be gained by actual measured value at
time
k
t
and predicted value at time
k
t
, it is simple iterative
relationship without complex operation; and the other is that it
takes advantages from full-term average as well as moving
average and doesnt abandon the old data. However, according
to how long the sampling time of data away from the present
moment, different weights are assigned to historical data and
the influence of historical data on the current forecast gradually
abates.
From the perspective of application, the most difficult point
of exponential smoothing prediction algorithm is how to
choose the parameter. On the one hand, the exponential
smoothing prediction algorithm is very convenient for use
because it need only one parameter; on the other hand, the
accuracy and reliability of prediction results are changed when
we set different value for the smooth parameter. In order to
improve quality of prediction algorithm and prediction result,
people did lots of explorations and researches from different
angles and suggested many kinds of improved algorithms in the
past half of the century. For example, some modification
algorithms were suggested, such as the double, triple and
multiple exponential smoothing prediction algorithms [1]. In
recent years, people come up with a variety of adaptive
algorithms and selection criteria [3-5] so as to find out some
available selection method about smoothing parameter. In
order to establish a new collaborative forecasting model,
literature [6] also combined the exponential smoothing
forecasting algorithm with neural network.
What is more, a lot of practice in using the exponential
smoothing prediction show that the exponential smoothing
prediction algorithm is lack of outlier-tolerant ability [7-8],
because the exponential smoothing prediction with constant
parameter is virtually linear prediction algorithm. In fact, the
prediction results would inevitably produce serious distortion
due to the impact of outliers when there are outliers in the
sample sequences.
So, in order to improve the prediction quality and to make
sure reliability of prediction results, there are two "bottleneck"
problems which are open: one is how to select adaptively the
smoothing parameter and the other is how to prevent bad
Vol. 2, No. 11, 2011
131 | P a g e
influence of outliers on the prediction algorithm. In this article
we build firstly an useful online adaptive smoothing parameter
estimation and then make further outlier-tolerant modification
on prediction algorithm. At the end of this paper, a practical
application is given to show that new algorithms are available.
II. ADAPTIVE FAULT-TOLERANT DESIGN OF
PARAMETERS
The exponential smoothing prediction algorithm is based on
the actual value
) (
k
t y
of the dynamic time-series in the
k
t

period and prediction value
) (
k
t y
predicted at time
k
t
.The
equations to calculate an exponential smoothing algorithm can
be expressed as follows:
=
+ =
+
) ( ) (
) ( ) ( ) 1 ( ) (
1 1
1
t y t y
t y t y t y
k k k
o o
(1)
where is a smoothing constant parameter, 1 0 s so .
How to select the parameter is important because
different values of may bring on different predicting value of
a dynamic process. For example, figure 1(a) is a series of
simulation sampling data, the solid line and the dotted line in
figure 1(b) are one-step prediction plots of figure 1(a) with two
different parameter (=0.2 and =0.8) respectively. Comparing
these two prediction curves shown in figure 1(b) clearly shown
that the value of has significant effect on smoothing
prediction results.
How to choose the smoothing parameter to get reliable
prediction effect? An intuitive idea is to seek such a parameter
estimator of which satisfying the following relation:
min )) ( ) ( ( ) (
1
2
o
o =
=
k
i
i i k
t y t y S
(2)
By the recursive formula (1), right side of the expression
(2) can be written as follows
=

+ =
k
i
i i i i k
t y t y t y t y S
1
2
1 1 1
)]} ( ) ( [ )] ( ) ( {[ ) ( o o

then the minimum point series of the parameter in the
function
) (o
k
S
can be obtained as follows:

=

=

=
k
i
i i
k
i
i i i i
k
t y t y
t y t y t y t y
1
2
1 1
1
1 1 1
) 0 (
)] ( ) ( [
)] ( ) ( )][ ( ) ( [
o
(3)
In order to ensure that design value of the smoothing
parameter meets the constraint condition that the parameter is
a nonnegative real value and
1 0 s so
, we can change
formula (3) into formula (4):
}} 1 , min{ , 0 max{
) 0 (
k k
o o = (4)

(a) Smulation data plot

(b) Exponential smoothing prediction value plot
Figure 1. Simulation and exponential prediction plots
We can get online optimal estimation of smoothing
coefficient from formula (3). Therefore, taking formula (1) and
(3) into consideration, we can build a new form of adaptive
exponential smoothing prediction algorithm as follows

)) ( ) ( (
)] ( ) ( [
)] ( ) ( )][ ( ) ( [
) ( ) (
1
2
1 1
1
1 1 1
1 k k
k
i
i i
k
i
i i i i
k k
t y t y
t y t y
t y t y t y t y
t y t y

+ =
=

=

+

(5)
with the beginning value ) ( ) (
1 1
t y t y = .
Using the formula (4), if quality of sampling sequence data
is reliable, the ordinary exponential smoothing prediction
algorithm (1) is changed into a new adaptive prediction
algorithm (5), which can be used effectively to solve the open
problem state in section I: how to choose the smooth
coefficient and ensure that the forecast results can effectively
track and early show the change of process.
The adaptive exponential smoothing prediction algorithm
(5) is concise and practical. Whats more, it is easy and
convenient to choose the smoothing coefficient parameter.
From engineering data processing and the actual process
prediction viewpoint, the algorithm (5) as well as algorithm (1)
does not have the ability of outlier-tolerance for isolated
outliers and patchy outliers: when the sampling data series of a

-2
-1
0
1
2
1 21 41 61 81 101 121
n(nubmer of data)
s
i
m
u
l
a
t
i
o
n

v
a
l
u
e

y

-2
-1
0
1
2
1 21 41 61 81 101 121
n(number of data)
p
r
e
d
i
c
t
i
o
n

v
a
l
u
e

y
parameter a=0.2
parameter a=0.8
Vol. 2, No. 11, 2011
132 | P a g e
dynamic process seriously deviate from the real state trend in a
time or time period, subsequent predictions values of local arcs
will distort or cause a big prediction deviation. So we will build
a group of outlier-tolerant exponential smoothing prediction
algorithms to improve outliertolerance ability of prediction
algorithm for outliers and ensure that the predictions will not
cause deviation even there are outliers occasionally in the
sampling.
Following the famous research ideas about robust statistics
suggested by Huber, we set the appropriate threshold parameter
(c,c1,c2,c3) and select Huber-type

function [9] and the re-
descending (short as Rd-)-type

function [10] in figure 2,
then the formula (2) can be written as follows:
min ) )) ( ) ( (( ) | (
1
2
o
o =
=
k
i
i i k
t y t y S

(6)
X
-c c
X
X
(x)
(x) (x)
2 -
(a) Huber
(b) Hample (c ) Rd-
-c c -c c c1 -c1 -c2 c2 -c3 c3

(a) Huber-type function (b) Rd-type function
Figure 2. Two kinds of function with fault-tolerance
Similarly, we also can get an outliertolerant algorithm of
for the exponential smoothing prediction parameter :
=

=
=

=

}} 1 ), ( min{ , 0 max{ ) (
)) ( ) ( )( )) ( ) ( ((
)) ( ) ( ))( ( ) ( )( )) ( ) ( ((
) (
) 0 (
1
2
1 1
2
1
1 1 1
2
) 0 (
| o | o
|
|
| o
k k
k
i
i i i i
k
i
i i i i i i
k
t y t y t y t y
t y t y t y t y t y t y
(7)
Accordingly, algorithm (5) and algorithm (1) can be
amended as the following form of adaptive fault-tolerant
exponential smoothing prediction algorithm as follows:
=
+ =
+
) ( ) (
)) ( ) ( ( )) ( 1 ( ) ( ) (
1 1
1
t y t y
t y t y t y t y
k k k k k
| | o (8)
In formula (8), the
|
function
) ( ) ( x x | & =
. If the

function is the Huber-type function or Rd-type function in
figure 2, then the
|
function is shown in figure 3.
X
-c c
X X
3
(a) Huber
(b) Hample
(c ) Rd-
-c c -c c
c1
-c1 -c2
c2
-c3
c3
X
) (x | ) (x |
) (x |
|

(a) Huber type
|
function (b) Rd-type
|
function
Figure 3. Two kinds of | function with fault-tolerance
Intuitively, if the
|
function is from figure 3, then the
exponential smoothing prediction algorithm (8) is outlier-
tolerant. In fact, it can be seen from the figure 3 that so long as
the forecast residual
) ( ) (
~
) (
~
k k k
t y t y t y =
is below the
reasonable scope threshold, two kinds of
|
function can
make full use of innovation brought by new samples; However,
once one-step predicting residual is beyond fixed threshold, the
Huber-type
|
function has the ability to directly eliminate
some negative influence from abnormal innovation on the of
subsequent prediction; the Rd-type
|
function can make full
use of the advantages of innovation and gradually reduce its
impact on predicting results, according to the unusual degree
with innovation.
III. ACTUAL DATA BASED CALCULATION AND
ANALYSIS
Using data series as shown in figure 4(a) and the adaptive
exponential smoothing prediction algorithm (5), the calculation
results were shown in figure 4(b). Comparing the figure 4(b)
with the figure 4(a), we may find out that the algorithm (5) can
well forecast process variations. However, it is sensitive to
outliers emerged in the sampling process, even causing a partial
distortion to prediction results.
Using the formula (7)~(8) of the adaptive outlier-tolerant
exponential smoothing prediction algorithm, and choosing the
Rd-type
|
function
) 7 , 5 , 3 (
1 1 1
o o o = = = c c c
,we get one-step
ahead prediction plot shown in figure 4(c). As can be clearly
seen from these calculation data sequence, using the adaptive
fault-tolerant exponential smoothing prediction algorithm can
effectively avoid the adverse impact on outliers in sampling
sequence and accurately predict the change of process status.
IV. CONCLUSION
Comparing figure 4(c) with figure 4(a),it is shown that the
adaptive outlier-tolerant exponential smoothing prediction
algorithm has the ability to forecast process variations
accurately under normal conditions for data sampling. Even if
there are a few outliers in sampling data, the adaptive outlier-
tolerant exponential smoothing prediction algorithm can also
do well.
Vol. 2, No. 11, 2011
133 | P a g e

(a) Actual data plot

(b) Adaptive exponential smoothing prediction plot

(c) Adaptive fault-tolerant exponential smoothing prediction plot
Figure 4. Adaptive exponential smoothing prediction and adaptive fault-tolera
nt exponential smoothing prediction curves of spacecraft sensor temperature

The research contents of this paper and the results obtained
have widely application value. The adaptive outlier-tolerant
exponential smoothing prediction algorithm can be widely used
in many different fields, such as space control and process
monitoring and economic forecasting and sensor fault
detection.
In fact, a lot of change in dynamic process can be
transformed into abnormal changes in measurement data or in
system state.
ACKNOWLEDGMENT
The research was supported by the National Nature Science
Fund of China (No.61074077) and the Tianyuan Fund of
National Nature Fund of China (No.11026224).
REFERENCES
[1] Gardner ES Jr. 1985. Exponential smoothing: the state of the art. Journal
of Forecasting 4: 1-28, 1985.
[2] Xiangbao Gao, Hanqing Dong. Data analysis and application of SPSS,
Beijing: Tsinghua University Press, 2009.4.
[3] R Snyder and J Ord. Exponential smoothing and the Akaike information
criterion. Monash University,2009.6
[4] S Makridakis,S Wheelwright and R Hyndman. Forecasting: methods and
application. New York, John Wiley & Sons. 1998
[5] Changjiang Wang. Research on smooth coefficient choice with
exponential smoothing method. Journal of North University of China,
vol .6:6-12,2006
[6] Kin Keung Lai, Lean Yu and Shouyang Wang. Hybridizing exponential
smoothing and neural network for Financial Time Series Predication.
V.N. Alexandrov et al. (Eds.): ICCS 2006, Part IV, LNCS 3994, pp. 493
500, Springer-Verlag Berlin Heidelberg,2006.
[7] Hu Shaolin, Sun Guoji. Process monitoring technology and application
(in Chinese).Beijing:National Defence Industry Press ,2001
[8] Hu Shaolin,Wang Xiaofeng,Karl Meinke,Huajiang Ouyang.Outlier-
tolerant fitting and online diagnosis of outliers in dynamic process
sampling data series.in Artificial Intelligence andComputational
Intelligence, pp.195-204, LNAI 7004(Deng Miao,Lei Wang,eds)
Springer Press,2011
[9] P J Huber. Robust statistics,John &Sons Press,1981
[10] Fan Jincheng,Hu Feng. Minimax robustly redescending estimators for
multivariate location vector.Journal of Xian Jiaotong University, vol.
29,no.12,pp:107-112,1995
AUTHORS PROFILE
Dr. Hu Shaolin graduated from Xian Jiaotong University of China and
received his PhD degree in Control Science and Engineering in 2000. He
pursued his postdoctoral studies in University of Science and Technology of
China in 2005. Currently, Dr. Hu Shaolin is a professor and researcher at the
State Key Laboratory of Astronautics. His research interests and publications
are in Control Engineering, System Safety, Digital Signal Processing, and
Data Mining.
Miss Li Ye graduated from Zhongbei University of China and received her
bachelor's in Automation in 2010. Currently, Miss Liye is a postgraduated
student at Xian University of Technology of China. Her research interests are
in navigation and digital signal processing.
Mr Zhang Wei graduated from Southeast University of China and received
her masters degree in Automation in 2006. Currently, Mr Zhangwei is an
engineer at the State Key Laboratory of Astronautics..His research interests
are in digital signal processing.
Mr Fan Shunxi graduated from Henan Normal University of China and
received his bachelor's degree in Automation in 2009. Currently, Mr Fan is a
postgraduated student at Xian University of Technology of China. His
research interests are in outlier-tolerant Kalman filter with applications in
navigation and digital signal processing.

10
20
30
40
50
0 200 400 600 800
t(h)
t
e
m
p
e
r
a
t
u
r
e
(
)
10
20
30
40
50
0 200 400 600 800
t(h)
t
e
m
p
e
r
a
t
u
r
e
(
)
10
20
30
40
50
0 200 400 600 800
t(h)
t
e
m
p
e
r
a
t
u
r
e
(
)
Vol. 2, No. 11, 2011
134 | P a g e
Towards Quranic reader controlled by speech

Yacine Yekache, Yekhlef Mekelleche, Belkacem Kouninef
Institut National des Tlcommunications et des TIC
BP 1518 Oran El Mnaouer 31000 Oran, ALGERIA

AbstractIn this paper we describe the process of designing a
task-oriented continuous speech recognition system for Arabic,
based on CMU Sphinx4, to be used in the voice interface of
Quranic reader.
The concept of the Quranic reader controlled by speech is
presented, the collection of the corpus and creation of acoustic
model are described in detail taking into account a specificities of
Arabic language and the desired application.
Keywords-arabic speech recognition; quranic reader; speech
corpus; HMM; acoustic model.
I. INTRODUCTION
Automatic speech recognition (ASR) is the technology that
permits the communication with machine using speech. There
are several applications that use this technology such as hands-
free operation and control as in cars or person with disabilities,
automatic dictation, government information systems,
automatic query answering, telephone communication with
information systems, etc.
The most dominant approach for ASR system is the
statistical approach Hidden Markov Model(HMM), trained on
corpora that contain speech resource from a large number of
speakers to achieve acceptable performance; unfortunately
there is a lack of this corpus for Arabic language. In this work
we collected a new corpus called Quranic reader command and
control which we will use to create an acoustic model using
sphinx train
II. QURANIC READER AND ARABIC LANGUAGE
A. Quranic reader
Quran is the central religious text of Islam, which is the
verbatim word of God and the Final Testament, following
the Old and New Testaments. It is regarded widely as the
finest piece of literature in the Arabic language. The Quran
consists of 114 chapters of varying lengths, each known as
a sura. Chapters are classed as Meccan or Medinan, depending
on when (before or after Hijra) the verses were revealed.
Chapter titles are derived from a name or quality discussed in
the text, or from the first letters or words of the sura.
There is a crosscutting division into 30 parts of roughly
equal division, ajza, each containing two units called ahzab,
each of which is divided into four parts (rub 'al-ahzab).
The Quran is the muslims way of life and the guidance
from Allah for that every muslim should read, listen and
memorize it; nowadays there are computer tools used for this
purpose, the interaction with this tools is by using a mouse or a
keyboard but in some situation it is difficult to use them for
example when driving a car or for blind person; so our goal is
to create a Quranic reader controlled by speech
B. Arabic language
Quran is revealed in Arabic for that it is the official
language of 23 countries and has many different,
geographically distributed spoken varieties, some of which are
mutually unintelligible. Modern Standard Arabic (MSA) is
widely taught in schools, universities, and used in workplaces,
government and the media.
Standard Arabic has basically 34 phonemes, of which six
are vowels, and 28 are consonants. A phoneme is the smallest
element of speech units that makes a difference in the meaning
of a word, or a sentence. The correspondence between writing
and pronunciation in MSA falls somewhere between that of
languages such as Spanish and Finnish, which have an almost
one-to-one mapping between letters and sounds ,and languages
such as English and French, which exhibit a more complex
letter-to-sound mapping[1] .
Figure 1. letter to sound mapping in Arabic

(1)
Vol. 2, No. 11, 2011
135 | P a g e
III. STATISTICAL APPROACH FOR SPEECH RECOGNITION
A. Mathematical formulation of ASR problem
A block diagram of this fundamental approach to speech
recognition is given in Fig.2, which shows a sentence W being
converted to a speech signal s[n] via the speech production
process. The speech signal is then spectrally analyzed (by the
acoustic processor) giving the spectral representation, X = (X1,
X2, . . . , XL) for the L frames of the speech signal. The
linguistic decoder solves for the maximum likelihood sentence
that best matches X (i. e., maximizes the probability of W
given X) via the Bayesian formulation:

(2)

(3)

The maximization of (1) is converted to (2) using the Bayes
rule, and since the denominator term P(X) is independent of W,
it can be removed leading to the three-step solution of (3). Here
we explicitly denote the acoustic model by labeling P(X|W) as
PA(X|W), where A denotes the set of acoustic models of the
speech units used in the recognizer, and we denote P(W) as
PL (W) for the language model describing the probabilities of
various word combinations. The process of determining the
maximum-likelihood solution is to first train (offline) the set of
acoustic models so that step 1 in (3) can be evaluated for each
speech utterance. The next step is to train the language model
for step 2, so that the probability of every word sequence that
forms a valid sentence in the language model can be evaluated.
Finally step 3 is the heart of the computation, namely a search
through all possible combinations of words in the language
model to determine the word [2]

Figure 2. Architecture of an HMM based recognizer
B. CMU Sphinx
Sphinx4 is a software implementation of HMM speech
recognizer, its architecture is highly flexible, Each labelled
element in Figure (3) represents a module that can be easily
replaced, allowing researchers to experiment with different
module implementations without needing to modify other
portions of the system. The main blocks in Sphinx-4
architecture are frontend, decoder and Linguist [3]
Front End: it parameterizes an input signal into a sequence
of output features. It performs Digital Signal Processing on the
incoming data.
--Feature: The outputs of the front end are features, used for
decoding in the rest of the system.
Linguist: Or knowledge base, it provides the information
the decoder needs to do its job, this sub-system is where most
of the adjustments will be made in order to support Arabic
recognition it is made up of three modules which are:
--Acoustic Model: Contains a representation of a sound,
created by training using many acoustic data.
--Dictionary: It is responsible for determining how a word
is pronounced.
--Language Model: It contains a representation (often
statistical) of the probability of occurrence of words.
Search Graph: The graph structure produced by the
linguist according to certain criteria (e.g., the grammar), using
knowledge from the dictionary, the acoustic model, and the
language model.
Decoder: It reads features from the front end, couples this
with data from the knowledge base and feedback from the
application, and performs a search to determine the most likely
sequences of words that could be represented by a series of
features.
The Configuration Manager: This sub-system is designed
in order to make Sphinx-4 pluggable and flexible.
Configuration information is provided in XML files, and the
actual code uses the Configuration Manager to look up the
different modules and sub-systems.

Figure 3. Sphinx-4 architecture
The Sphinx framework was originally designed for English,
but nowadays it supports also, among others, Spanish, French
and Mandarin. However, there are limited acoustic model
available for Arabic [4, 5, 6].
Vol. 2, No. 11, 2011
136 | P a g e
Our goal is to use Sphinx to build Quranic reader controlled
by speech ; the main tuning that we should do in the frame
work is the creation of the acoustic model and the application
that interact with the framework. In the following section we
will describe the creation of the acoustic model.
IV. ACOUSTIC MODELING
The creation of acoustic model pass through two steps, the
first is data collection (corpus) and the second is training the
model using this data,
A. Data collection (corpus )
The following elements were required to train a new acoustic
model:
Audio data with recorded speech;
Transcription of each audio file;
Dictionary with phonetic representations of all words
appearing in the transcriptions;
List of phonemes (and sounds) appearing in the
transcriptions
B. Speech collection
We prepared a text file which contain 114 suras names,
famous receiters names and control words that must be
correctly recognized for the Quranic reader to function
properly, The amount of audio data required to properly train
the model depends on the type of the model. For a simple
command-and-control one-speaker system the amount of data
can be fairly low. For multi-speaker systems the amount of
required audio increases and this is our case.
After selecting text for the recognizer, recording of this
chosen data is required. For this work, the recording has been
taken place using 50 speakers from different region of Algeria;
their personal profile includes information like name, gender,
age, and other information like Environmental condition of
recording (for example: class room condition, sources of noise
like fan, generator sound etc) Technical details of device (pc,
microphone specification)
The audio file was recorded using sampling rate of 16KHZ
and 16 bit per sample, this rate was chosen because it provide
more accurate high frequency information, after that the
splitting by command and word was done manually and saved
in .wav format[7]. Each file has been named using this
convention: speakername-commandID.wav for example a file
with the name yacine-001 mean that this file is recorded by
yacine and it contain the Fatiha word
These audio files was divided into two sets, the first is
composed of data from 20 males and 15 female used to train
the acoustic model and the second composed of data from 10
males and 5 female for testing purpose.
C. Transcription file
The second step is the transcription of the training set of the
collected audio files; any error in the transcription will mislead
the training process later. The transcription process is done
manually, that is, we listen to the recording then we match
exactly what we hear into text even the silence or the noise
should be represented in the transcription.
D. Phonetic dictionary
In this step we mapped each word in the vocabulary to a
sequence of sound units representing pronunciation; that it
contained all words with all possible variants of their
pronunciation, to take into account pronunciation variability,
caused by various speaking manners and the specificity of
Arabic. . Careful preparation of phonetic dictionary prevents
from incorrect association of a phoneme with audio parameters
of a different phoneme which would effect in decreasing the
models accuracy.
For example

E. List of phoneme
This is a file which contain all the acoustic units that we
want to train model for, The SPHINX-4 does not permit us to
have units other than those in our dictionaries. All units in the
dictionary must be listed here. In other words, phone list must
have exactly the same units used in your dictionaries, no more
and no less. The file has one phone in each line, no duplicity is
allowed.
TABLE1. WORDS USED IN THE CORPUS
Sura name

........
001
002
003
004
114
Reciters names

115
116
117
118

Control

119
121
121
122
123
124
125
126
127
128
129
131
131
132
Arabic digit

..

133
134
....
142
F. Acoustic model training
Before acoustic modeling we should extract features
vectors from the speech for a purpose of training. The
Vol. 2, No. 11, 2011
137 | P a g e
dominating feature extraction technique known as Mel-
Frequency Cepstral Coefficients (MFCC) was applied to
extract features from the set of spoken utterances. A feature
vector Y represents unique characteristics of each recorded
utterance,
The most widely used method of building acoustic models
is HMMs, Each base phone q is represented by a continuous
density HMM of the form illustrated in Fig.(4) with transition
parameters {aij } and output observation distributions
{bj ()}.entry and exit states are nonemitting and they are
included to simplify the process of concatenating phone models
to make words.

Figure 4. HMM based phone model
Each word in our corpus should be modeled using
HMM,the parameter aij and bj are estimated from our
collected corpus using expectation maximization (EM). For
each utterance, the sequence of base forms is found and the
corresponding composite HMM constructed.

Figure 5. The BaumWelch training procedure based on a given training set
of utterances
A forwardbackward alignment is used to compute state
occupation probabilities (the E step) and the means and
covariances are then estimated via simple weighted
averages(theMstep) [8]
To create the acoustic model we used sphinx train, which need
as input the recorded speech, transcription, dictionary and
phoneme files to produce the acoustic model. Much of Sphinx
Train's configuration and setup uses scripts written in the
programming language Perl.
V. CONCLUSION
In this paper we reported the first steps toward developing
Quranic reader controlled by speech using sphinx4 framework,
in this steps we specified the words that should be recognized
and collected a corpus used to train the acoustic model with
sphinx train.
Further we will integrate the acoustic model in sphinx 4 and
build an application that interacts with sphinx framework.
REFERENCES
[1] Nizar Y. Habash, Introduction to arabic natural language processing
2010 by Morgan & Claypool.
[2] J. L.R. Rabiner, A tutorial on hidden Markov models and selected
applications in speech recognition,Proc. IEEE 77(2), 257286 (1989).
[3] Willie Walker, Paul Lamere, Philip Kwok, Bhiksha Raj, Rita Singh,
Evandro Gouvea, Peter Wolf, Joe Woelfel, Sphinx-4: A Flexible Open
Source Framework for Speech Recognition,. SMLI TR2004-0811
c2004 SUN MICROSYSTEMS INC
[4] H. Hyassat, and R. Abu Zitar, "Arabic speech recognitionusing SPHINX
engine," International Journal of SpeechTechnology, Springer, pp. 133-
150,2008.
[5] K. Kirchhoff, J. Bilmes, S. Das, et aI., "Novel approaches to Arabic
speech recognition: report from the 2002 Johns-Hopkins summer
workshop," ICASSP'03, Hong Kong, vol.l, pp. 344-347,2003
[6] M. Ali, M. Elshafei, M. Alghamdi, H. Almuhtaseb, and A. Al-Najjar,
Generation of Arabic Phonetic Dictionaries for Speech Recognition,
IEEE Proceedings of the International Conference on Innovations in
Information Technology, UAE, pp. 59 63, 2008.
[7] Artur Janicki , Dariusz Wawer Automatic Speech Recognition for
Polish In a Computer Game Interface Proceedings of the Federated
Conference on Computer Science and Information Systems pp. 711716
[8] Jacob Benesty, M. Mohan Sondhi, Yiteng Huang Springer handbook of
speech processing partE p 554
[9] Alghamdi, Mansour, Fayez Alhargan, Mohamed Alkanhal,Ashraf
Alkhairy, Munir Eldesouki and Ammar Alenazi1 (2008) Saudi Accented
Arabic Voice Bank. Workshop on.Experimental Linguistics. Athens,
Greece
[10] Artur Janicki , Dariusz Wawer Automatic Speech Recognition for
Polish In a Computer Game Interface Proceedings of the Federated
Conference on Computer Science and Information Systems pp. 711716
[11] Jacob Benesty, M. Mohan Sondhi, Yiteng Huang Springer handbook of
speech processing partE p 554.
[12] Introduction to Arabic Speech Recognition Using CMUSphinx
System by H Satori, M Harti, N Chenfour

Vol. 2, No. 11, 2011
138 | P a g e
Design and Implementation for Multi-Level Cell
Flash Memory Storage Systems
Amarnath Gaini, K Vijayalaxmi
Assistant Professor
Department of Electronics
VITS (N9), Andhra Pradesh
Sathish Mothe
Assistant Professor
Department of Electronics
JITS (27), Andhra Pradesh

Abstract The flash memory management functions of write
coalescing, space management, logical-to-physical mapping, wear
leveling, and garbage collection require significant on-going
computation and data movement. MLC flash memory also
introduces new challenges: (1) Pages in a block must be written
sequentially. (2) Information to indicate a page being obsolete
cannot be recorded in its spare area. This paper designs an MLC
Flash Translation Layer (MFTL) for flash-memory storage
systems which takes new constraints of MLC flash memory and
access behaviors of file system into consideration. A series of
trace driven simulations is conducted to evaluate the
performance of the proposed scheme. Our experiment results
show that the proposed MFTL outperforms other related works
in terms of the number of extra page writes, the number of total
block erasures, and the memory requirement for the
management.
Keywords-Flash memory; MFTL; MLC; BAST; FAST.
I. INTRODUCTION
Flash memory chips are constructed from different types of
cells (NOR and NAND), and with different numbers of cells
per memory location (single-level cell or SLC; and multi-level
cell or MLC). These variations result in very different
performance, cost, and reliability characteristics. NOR flash
memory chips have much lower density, much lower
bandwidth, much longer write and erase latencies, and much
higher cost than NAND flash memory chips. For these reasons,
NOR flash has minimal penetration in enterprise deployments;
it is primarily used in consumer devices. Leading enterprise
solid state drives (SSDs) are all designed with NAND flash.
Another distinction in flash memory is SLC versus MLC.
MLC increases density by storing more than a single bit per
memory cell. With their increased density, the cost of MLC
flash chips is roughly half that of SLC, but the MLC write
bandwidth is about 2 times worse than SLC, and MLC supports
from 3 to 30 times fewer erase cycles than SLC. A new
generation of SSDs incorporates special firmware that closes
the performance and durability gap between SLC and MLC.
Flash-memory storage systems are normally organized
inlayers, as shown in Figure. 1.
The Memory Technology Device (MTD) layer provides
lower-level functionalities of flash memory, such as read,
write, and erase. Based on these services, higher-level
management algorithms, such as wear-leveling, garbage
collections, and logical/physical address translation, are
implemented in the Flash Translation Layer (FTL). The
objective of the FTL is to provide transparent services for file
systems such that flash memory can be accessed as a block
oriented device.
As an alternative approach, a file system can be made flash-
memory-aware by having the characteristics of flash memory
been taken into consideration. JFFS and YAFFS take this
approach. Although a flash-memory aware file system is more
efficient, FTL scheme has the advantage of enabling flash
memory available to existing file systems directly. Since most
of the applications, especially for portable devices, access their
files under FAT file systems, this paper focuses on FTL.

Figure 1. Architecture of Flash-Memory Storage Systems
Depending on the granularity with which the mapping
information is managed, FTL can be classified into page-level,
block-level, and hybrid mapping schemes. For a page-level
mapping scheme, each entry in the mapping table includes a
physical page number (PPN), which is indexed by the logical
page number (LPN).
When a request arrives, the mapping scheme translates the
associated logical address into an LPN and uses this LPN to
locate the corresponding PPN from the mapping table. With the
mapping table, a logical address can be directly mapped to a
page in any location of a flash memory. It makes storage
management more flexible and results in better performance
for random accesses. Since the mapping table is large in size
Vol. 2, No. 11, 2011
139 | P a g e
and usually resides in SRAM, a page-level mapping scheme is
considered impractical due to cost consideration.
To reduce SRAM requirement, block-level mapping
scheme was proposed. In a block-level mapping scheme, a
logical address has two components: a logical block address
(LBA) and a logical page offset. LBA can be translated into a
physical block address (PBA), and the logical page offset is
used to locate the target page in the physical block. The size of
the mapping table is now significantly reduced and
proportional to the total number of blocks in flash memory.
However, update requests usually incur a block-level copy
overhead. Block level mapping shown in fig 2.

Figure 2: Block level mapping
A hybrid mapping scheme is a block-level mapping scheme
with a limited number of records adopting page-level mapping.
In a hybrid scheme, a logical address is consisted of an LBA
and a logical page offset as a block-level scheme. However,
unlike the block-level scheme, it allows data update to the next
available page in another physical block, i.e., a log block. A
page mapping table is required for recording the actual
locations of the updated data. Hybrid mapping schemes
consume an extra amount of SRAM space for the records with
page-level mapping but reduces the copying overhead of a
block-level mapping scheme.
Block Associative Sector Translation (BAST) is a hybrid
mapping scheme proposed by Kim et al. [4]. BAST maintains a
limited number of log blocks to handle update requests. Each
data block may associate with at most one log block. When
processing an update request, the associated data is appended
to free pages of the corresponding log block sequentially, as
long as such free pages are available. Merge operation is
required when no free log block available. Fully Associative
Sector Translation (FAST) is another hybrid mapping scheme
proposed by Lee et al. [6]. FAST shares the log blocks among
all of the data blocks. It allows an update request to be written
to the current log block. As a result, FAST improves the
utilization of log blocks and reduces the number of merge
operations. Since a log block may associate with several data
blocks at the same time, merging a log block may require the
merging of several data blocks. Both BAST and FAST are
designed for single-level cell (SLC) flash memory.
Currently, multi-level cell (MLC) flash memory has
occupied the largest part of the flash memory market, due to an
increasing demand for a larger capacity and smaller size
storage medium. In MLC flash memory, each cell can store
two or more bits of data. Although MLC flash memory offers a
lower cost and higher density solution, it imposes additional
stringent constraints [8]. First, the number of partial program
cycles in the same page is limited to one. This constraint
implies that it is no longer permitted to mark a page as dead by
simply clearing some specific bit in the spare area, which is
widely adopted in previous researches. Another constraint is
related to the order in which write operations are performed. In
MLC, free pages of a block must be written consecutively from
the first free page to the target page in the block; random
written order is prohibited. Such a constraint makes most
existing block-level mapping schemes and hybrid mapping
schemes inapplicable for modern flash memory chips. Further,
direct modification to these mapping schemes is not
appropriate as it will inevitably lead to a significant increase of
overhead.
Motivated by the above concerns, we propose a hybrid
mapping scheme for the management of the modern MLC flash
memory. Since FAT file system is most commonly used for
accessing flash memory, the proposed scheme focuses on the
file access behavior under FAT file system as well. In the
proposed scheme, the mapping granularity would adjust
adaptively according to the access pattern. Our scheme is
distinguished from existing solutions in that the proposed
scheme consumes less SRAM while provides a better
performance. In addition, existing solutions does not take
stringent constraints of MLC flash memory into account while
our scheme could apply to SLC flash memory as well. The rest
of the paper is organized as follows. Section II presents the
proposed scheme in details. Section III evaluates the
performance of the proposed scheme and compares it with
existing works. Section IV draws the conclusion.
II. MFTL:MLC FLASH TRANSLATION
A. LAYER
1) Overview
When a write request arrives, MFTL calculates the
corresponding virtual block address (VBA) via the logical
address of the request, allocates a free (physical) flash-memory
block for the virtual block, and writes data to the allocated
block.
The PBA of the allocated block is recorded in Block Table
[VBA]. Note that VBA is similar to LBA in block-level
mapping schemes, and we use different notation to avoid
confusion. Since out-place update is a widely adopted solution
for handling data-update requests, an extra replacement block
(physical flash-memory block) is assigned to store the newly
updated data. To keep track of the latest version of data, an
UpdateRec is created whenever a replacement blocks is
allocated for a virtual block. Note that we do not maintain
UpdateRec for every virtual block all the time as we do for
Block Table. In fact, as shown later in the experiments
maintaining five UpdateRecs would be enough to deal with
sustained read/write requests and MFTL always keeps the
latest UpdateRecs in SRAM. Since data updates might be
random and data amount of each update might be small, only
UpdateRec is insufficient if fast data access is required.
2) Data Structure

Each UpdateRec contains the following fields:
Vol. 2, No. 11, 2011
140 | P a g e
a) VBA records the corresponding virtual blockaddress.
b) Primary records the physical block address of the
associatedprimary block.
c) Replace records the physical block address of The
associatedreplacement block.
d) LWP is the index of the last written page in the
associatedreplacement block.
e) Count maintains the number of backward updates
occurred in thisrecord. The initial value of Count is
0.
f) Mode is a one-bit flag indicating whether the
mapping isin block mapping mode (0) or
pagemapping mode (1).
g) Priority is the priority for UpdateRec replacement.
Whenwrite/update request arrives, Priority of The
correspondingUpdateRec (if any) is set to 0xFF,
while Priority ofotherUpdateRecs would be
decreased by 1.

Page Map Status is introduced to maintain page-mapping
information for virtual blocks. When an UpdateRec switches
from block mapping mode to page mapping mode, MFTL has
to merge the corresponding primary block into the replacement
block and then treat the resulting replacement block as the
primary block. We can eliminate such overhead by directly
switching to page mapping mode without merging operation
when LWP is less than half of total pages in a block. The most
significant bit (MSB) is used to indicate whether the mapped
age is located in the primary block (set to 0) or in the
replacement block (set to 1). Each Page Map Status contains
two elements:
a) VBA records the corresponding virtual blockaddress.
b) MapArray[N] is an integer array keeping the page
mapping information of the block,

Where N is the number of pages per block. A value smaller
than 0x80 in the Map Array[i] means the i-th logical page is
located at the Map Array[i]-th page of the primary block.
Otherwise, the i-th logical page is located at the (Map Array[i]
AND 0x7F)-th page of the replacement block.

3) Write Flow

When the FAT file system creates a new file, it first issues
a write to root directory to update the file information. It then
updates FAT tables to allocate required storage space for the
file. Finally, the file body is written to the allocated area.
Conceptually, the storage space could be treated as containing
two parts: system area and data area. System area consists of a
boot record, FAT tables and a root directory, while data area is
the area used to store file body. Any write/update to a file (data
area) is accompanied by several small updates to the system
area to ensure correct file information. Observing the very
different behavior over system area and data area, we adopt
various mechanisms to deal with write operations in an
adaptive manner. Since writing data is trivial in flash-memory
management, we will focus on the operation of data updating.
Upon arrival of an update request, MFTL first determines
whether the corresponding UpdateRec exists by checking the
VBA field of each UpdateRec. If no such an UpdateRec
exists, MFTL allocates a new one.

In case that all the UpdateRecs have been assigned and the
new update request does not match any one of them, MFTL
will select the UpdateRec with the lowest priority as a victim
for replacement (ties are broken arbitrarily). Live pages in the
primary block and the replacement block of the victim
UpdateRec will be merged to a newly allocated block, and the
victim UpdateRec could then be released and reassigned to the
new request. After the UpdateRec is located, MFTL determines
the mapping mode of the UpdateRec. Different mapping mode
would affect the manner how MFTL writes data to flash
memory.
If the mapping mode is in block mapping mode, some
further checks are required. According to the index of starting
target page for the update request (i.e., STP) relative to the
index of the last written page in the replacement block (i.e.
LWP); the update request is treated differently. For the case in
which STP is larger than LWP, the request is treated as a
forward update. Otherwise, some previously updated data is to
be updated again by the request, and the request is treated as a
backward update. Since backward updates incur higher page
copying overheads, page-mapping mode is thus introduced to
deal with small, random, and frequently updated data.
The Count Field in UpdateRec is used to maintain the
number of backward updates that had occurred in this record.
When it exceeds a threshold, the mapping is switched to page
mapping mode. If a backward update occurs but does not result
in a switch to page mapping mode, MFTL will examine
whether page-copying after updating data is required. The
primary reason behind page-copying after updating data is the
assurance of data integrity, since the primary block or the
replacement block would be reassigned in this case.
4) Read Flow

For a read request, MFTL locates the target data through
either Block Table or UpdateRecs (and Page Map Statuss, if
required).When a read request arrives, MFTL first derives the
corresponding VBA and STP via the logical address of the
request. Then, MFTL determines whether the derived VBA
matches the VBA field of any UpdateRec. If no such
UpdateRec is found, the PBA of the target data can be obtained
from Block Table [VBA]. With the PBA, MFTL accesses the
corresponding flash-memory block and read the required data
from page STP to the last page of the block. Suppose that the
UpdateRec is successfully found, and it is in block mapping
mode. In this case, MFTL has to identify which block and how
many pages shall be read. It compares the derived STP with
LWP of the UpdateRec.
If STP is smaller than or equal to LWP, a portion of the
required data shall be read from the corresponding replacement
block. MFTL reads the required data from page STP to page
Vol. 2, No. 11, 2011
141 | P a g e
LWP in the replacement block and then read from page
LWP+1 to the last page in the primary block. Otherwise,
MFTL simply reads the required data from STP to the last page
in the primary block. For the case in which the corresponding
UpdateRec is found and in page mapping mode, MFTL may
not be able to read the required data sequentially because the
live pages in the replacement block are not programmed in
such a manner. MFTL must look up Map Array to determine
physical locations of the needed data. For each target page i, if
the value of Map Array[i] is 0xFF, MFTL reads data from page
i in the primary block. Otherwise, MFTL reads from page Map
Array[i] in the replacement block.
III. PERFORMANCE EVALUATION
A. Experiment Setup
This section is meant to evaluate the performance of the
proposed MFTL in terms of the number of extra page writes2,
the number of total block erasures, and the memory
requirement for the management. We compared the proposed
MFTL with two well-known hybrid mapping schemes, BAST
[4] and FAST [6], under different number of extra blocks.
Suppose the number of extra blocks is set to n, MFTL could
have up to n UpdateRecs. The maximum number of blocks
reserved for random accesses in BAST could be n, while FAST
would contain one sequential write (SW) log block and (n
1)random write (RW) log blocks. The experiments were
conducted by a trace-driven simulation.
Traces with various access behaviors were collected under
Windows XP with FAT16 file system over a 2GB flash
memory. The flash memory chip adopted in the experiment
was Samsung K9WAG08U1B 16G-bit SLC flash memory
each block contains 64 pages, and each page is of 2KB. All of
the traces were captured by the SD/MMC card protocol
analyzer VTE2100 [9]. Note that the FAT file system would
update the root directory and FAT tables frequently to maintain
the correct file information. The cluster size was defined
as16KB, where the cluster size was the minimum allocation
unit for a file. TABLE I summarizes characteristics of the
seven evaluation patterns.
B. Experiment Results
Fig. 3 compares the number of extra page writes and the
number of total block erasures of BAST, FAST, and the
proposed MFTL under different number of extra blocks over a
2GB flash memory. In addition to the extreme case (i.e. the
pattern 4KB-R), FAST had a good performance when the
number of extra blocks was very limited; three extra blocks
would be enough to make its performance stable. This was
because FAST shares its extra blocks for all logical blocks.
However, the proposed MFTL outperformed both BAST and
FAST for all patterns when four extra blocks could be
provided. Based on our experimentation, the performance
improvement was significant when the number of extra blocks
was between four and five. Such results were caused by two
reasons: (1) Since the FAT file system would update two FAT
tables, root directory, and data area while writing a file to
storage, four extra blocks could significantly improve the
performance area while writing a file to storage, four extra
blocks could significantly improve the performance. (2)
Windows XP writes files with size over64KB by two threads;
five extra blocks might be required for updating a large file. As
a result, over six extra blocks would not have an obvious
improvement for extra page writes and block erasures.
TABLE I: CHARACTERISTICS OF EVALUATION PATTERNS

For the pattern 4KB-R, as illustrated in Fig. 3(a) and Fig.
3(b), the number of extra blocks seemed to have no effect on
the performance improvement. It was because a tremendous
amount of random 4KB file updates would rapidly use up log
blocks. For BAST, each subsequent update would trigger a
merge operation, and each merge operation would incur
64extra page writes and 2 block erasures. For FAST, since a
log block was shared by all data blocks, merge operations
would not be triggered as frequent as BAST did. Recall that
each flash-memory block contains 64 2KB pages and the
cluster size was 16KB, each merge operation would incur 8
64 = 512 extra page writes and 9 block erasures (compared
with 8 merge operations in BAST with 512 extra page writes
and 16 block erasures). For MFTL, random 4KB file updates
would not switch UpdateRecs into page mapping mode, and
each 4KB file update would be treated as a forward update in
block mapping mode (please refer to Section II-C1). In
addition, updates to root directory and FAT tables could always
reside in page mapping mode thanks to Count and Priority in
UpdateRec. As a result, FAST and MFTL could have a better
performance in extra page writes and outperform BAST in
block erasures. For the pattern 64KB-S, as illustrated in Fig.
3(c) and Fig.3(d), there was no internal fragmentation issue and
readers might expect similar performances for the three hybrid
mapping schemes. However, the experiment result showed a
different story. For BAST and FAST, while there was no log
block assigned to the corresponding data block and page off set
of the write request did not align to the block boundary (i.e.
page offset 0), a log block would be assigned and updated data
would be written to the page 0 of the block. Even the following
write requests were sequential writes; BAST and FAST would
still incur extra page writes and block erasures during merge
operations. On the other hand, as mentioned in Section II-C,
MFTL uses Count in UpdateRec to prevent from switching to
page mapping mode immediately.
Vol. 2, No. 11, 2011
142 | P a g e
This mechanism helped UpdateRec to keep in block
mapping mode for sequential write requests. Thus merge
overheads could be reduced, and a better performance was
achieved. For patterns SUB and MP3, since large files led to
much more sequential write operations, the number of merge
operations was largely reduced.
As shown in Fig. 3(f) and Fig. 3(h),the improvement over
the reduction of block erasures was not as significant as
patterns 4KB-R and 64KB-S; MFTL could achieved a better
performance on block erasures and extra page writes compared
With the other two schemes for large file patterns while at least
four extra blocks were provided. Note that in the pattern MP3
with at least four extra blocks, MFTL outperformed BAST and
FAST in terms of extra page writes. It was because MFTL uses
Count in UpdateRec to prevent from switching to page
mapping mode immediately. With a small overhead of page
copying, this mechanism helped UpdateRec to keep in block
mapping mode. As a result, merge overhead could be reduced
when an update request dose not begins from page 0 of some
block.
For the memory requirements of the three hybrid mapping
schemes, since block mapping tables were the same for each
scheme, we focused on the memory requirement for managing
extra blocks. TABLE II lists the memory requirements for each
extra block and each Map Array under the three different
schemes. BAST reserved one more byte to store information
for the selection of a victim log block when a merge operation
is triggered. MFTL reserved three extra bytes for fields of
LWP, Count, Mode, and Priority. For page mapping, BAST
required an array for each log block to keep track of the page
mapping information. Since each block contained 64 pages,
64entries were required for the array and each entry was of one
byte. FAST required an array for each RW block. Since each
RW block was shared by all data blocks, Each entry of the
array required three bytes to maintain complete information.
MFTL did not reserve an array for each UpdateRec. Thus
MFTL required two more bytes to record the corresponding
VBA.

Figure 3: Performance Comparison of BAST/FAST/MFTL under Different Number of Extra Blocks.

TABLE II: SRAM REQUIREMENTS FOR DIFFERENT SCHEMES (UNIT: BYTES)

TABLE III lists the memory requirements for these schemes
with respect to different number of extra blocks. As shown in
the table, when the number of extra blocks is five, MFTL can
save about 31.55% memory space compared with BAST and
up to 69.55% compared with FAST.
These savings can be as high as 59.44% and 83.89% when
there are ten extra blocks provided by the system.
TABLE III: SRAM REQUIREMENTS FOR MANAGING UPDATE REQUESTS WITH
DIFFERENT NUMBER OF EXTRA BLOCKS. (UNIT: BYTES)

Vol. 2, No. 11, 2011
143 | P a g e
IV. CONCLUSION
Flash technology is constantly changing, providing faster
program and erase cycles, a bigger number of guaranteed
erase and re-program cycles and longer data retention. This
paper proposes a management scheme for MLC flash memory
storage systems.
Observing that most of existing user applications access
NAND flash memory under FAT file system, this paper takes
new constraints of MLC flash memory and access behaviors
of FAT file system into consideration. The proposed MFTL
scheme treats different access behaviors in an adaptive
manner. It adopts a hybrid mapping approach to balance
between performance and SRAM requirement. An extensive
set of experiments has been conducted to evaluate the
performance of the proposed MFTL.
Our experiment results show that the proposed scheme out
performs BAST and FAST in terms of extra page writes and
block erasures while SRAM requirement is largely reduced.
For the future work, we shall further take the crash recovery
and wear leveling issues into our design and survey other
newly announced hybrid mapping schemes, e.g., LAST and
KAST, for comparisons.
REFERENCES
[1] Journaling Flash File System (JFFS) and Journaling Flash File System
2 (JFFS2), http://sources.redhat.com/js2/js2-html/
[2] Understanding the Flash Translation Layer (FTL) Specification.
Technical report, Intel Corporation, Dec 1998.
[3] Aleph One Company. Yet another Flash Filing System.
[4] H. Cho, D. Shin, and Y. I. Eom. KAST: K-Associative Sector
Translation for NAND Flash Memory in Real-Time Systems. In
Design, Automation and Test in Europe (DATE), pages 507512, April
2009.
[5] J. Kim, J. M. Kim, S. H. Noh, S. L. Min, and Y. Cho. A Space -
Efficient Flash Translation Layer for Compact Flash Systems. In IEEE
Transactions on Consumer Electronics, Vol. 48, No. 2, pages 366375.
[6] S. Lee, D. Shin, Y.-J. Kim, and J. Kim. LAST: Locality-Aware Sector
Translation for NAND Flash Memory-Based Storage Systems. ACM
SIGOPS Operating Systems Review, 42:3642, October 2008.
[7] S.-W. Lee, D.-J. Park, T.-S. Chung, D.-H. Lee, S. Park, and H.-J. Song.
A Log Buffer-Based Flash Translation Layer Using Fully-Associative
Sector Translation. In ACM Transactions on embedded Computing
Systems, Vol. 6, No. 3, Article 18, July 2007.
[8] M-Systems. Flash-memory Translation Layer for NAND flash (NFTL),
1998.
[9] Samsung Electronics. Samsung K9LBG08U0M(v1.0) - 32Gb DDP
MLC Data Sheet.
[10] Test metrix Inc. VTE2100.
[11] D. Woodhouse. JFFS: The Journaling Flash File System. In Ottawa
Linux Symposium, 2001.
[12] A set-based mapping strategy for flash-memory reliability
enhancement. Design, Automation & Test in Europe Conference &
Exhibition, 2009.
[13] Program/erase selection for flash memory, Jerry A. Kreifels et al.
[14] A flash-memory based file system, TCON'95 Proceedings of the
USENIX 1995 Technical Conference Proceedings USENIX
Association Berkeley, CA, USA 1995
[15] Dr. John R Busch; Flash Memory Architecture Alternatives.
AUTHOR PROFILE
Amarnath Gaini received the Bachelors. degree in
Electronics and Communication Engineering from
JNTUH, Hyderabad, India in 2007 and Masters
degree in VLSI System Design from JNTUH,
Hyderabad, India in 2009. He is currently working as
an Assistant professor in Electronics and
Communication Engineering department in VITS
(N9), Andhra Pradesh, India. He has a Life Member
ship in ISTE. His research interests in VLSI,
Embedded systems and digital systems platform.
Vol. 2, No. 11, 2011
144 | P a g e
A novel approach for pre-processing of face
detection system based on HSV color space and
IWPT

Megha Gupta, Prof. Neetesh Gupta
Information and Technology department
Technocrats Institute of Technology
Bhopal, Madhya Pradesh, India

Abstract Face detection system is challenging area of research
in the field of security surveillance. Preprocessing of facial image
data is very important part of face detection system. Now days
various method of facial image data preprocessing are available,
but all these method are not up to mark. Now we have a novel
approach for preprocessing of face detection system based on
HSV color space and integer wavelet packet transform method.
And finally in this paper we have used LBP (local binary pattern)
and SVM classification process are used.
Keywords- facial image, HSV color space, IWPT, SVM.
I. INTRODUCTION
In todays world human face detection is a prelude required
step of face recognition systems as well as very important
process in security based applications. Human often uses faces
to recognize persons and advancement in computing capability
over the past few decades. The development of face detection
system is quite essential in a variety of application such as
robotics, security system, and intelligent human-computer
interfaces, etc. A number of face detection methods such as
those using Eigen-faces [1] and neural network [2], have been
developed. In these methods however a large amount of
computation is required, making the processing extremely time
consuming. Initial step of any processing of face is detecting
the location in images where faces are present. But face diction
from a single image is difficult because of variability in scale,
locations, pose, and color.
There are many methods for the detection of face; color is
an important feature of human face. Using skin color as feature
for extracting a face increases the detection rate. Face
recognition can be used for both verification and identification.
Verification is done by comparing the two images. If the two
images got matched means the input image is verified.
Identification is done by comparing the input image with more
than one image and finding the closest match of that input
image. The method presented in this paper consists of three
steps- skin reorganization, face reorganization and face
recognition. The innovation of the proposed method is using
skin reorganization filter as a preprocessing step for face
reorganization. Now in this paper we proposed new
preprocessing technique using integer wavelet packet transform
and HSV color space for better preprocessing of facial image
data. Integer wavelet packet transform function is a special
function of transform function from these function we
decompose the facial image data without loss of information.
And approach of this paper is HSV color space for better
intensity of facial image instead of RGB model. The remainder
of this paper is organized as follows. Section II briefly reviews
the literature. A new approach for preprocessing of facial
image data is explained in Sect. III. Finally, section IV presents
conclusions and future scope
II. LITERATURE REVIEW
By Mayank Vatsa, Richa Singh [1] proposed method for
facial image preprocessing as Age Transformation using
Mutual Information Registration the face images are pre-
processed and quality of the image is normalized, we minimize
the age difference between gallery and probe face images. One
way to address the challenge of facial aging is too regularly
update the database with recent images or templates. However,
this method is not feasible for applications such as border
control and homeland security, missing persons and criminal
investigations address this challenge, researchers have
proposed several age simulation and modeling techniques.
These technique model the facial growth that occurs over
period of time to minimize the difference between probe and
gallery images. Unlike, the conventional modeling approach,
we proposed mutual information registration based age
transformation algorithm to minimize the age difference
between gallery and probe images. Mutual information is a
concept from information theory in which statistical
dependence is measured between two random variables [3, 4].
Researchers in medical imaging have used mutual
information based registration algorithms to effectively fuse
images from different modalities such as CT and MRI [3], [4].
We have used this algorithm because there may be variations in
the quality of pre-processed scanned and digital face images,
and the registration algorithm should contend with these
variations. Age difference minimization using registration of
gallery and probe face images is described as follows:
Let F
G
and F
P
be the detected and quality enhanced gallery
and probe face images to be matched. Mutual information
between two face images can be represented as,
M (F
G
, F
P
) = H (F
G
) + H (F
P
) H (F
G
, F
P
) (1)
Vol. 2, No. 11, 2011
145 | P a g e
Where, H (.) is the entropy of the image and H (F
G
, F
P
) is
the joint entropy. Registering F
G
with respect to F
P
requires
maximization of mutual information between

F
G
and F
P,
thus
maximizing the entropy H (F
G
) and H (F
P
), and minimizing the
joint entropy H (F
G
, F
P
). Mutual information based registration
algorithms are sensitive to changes that occur in the
distributions as a result of difference in overlap regions. To
address this issue, Hill et al. [3] proposed normalized mutual
information that can be represented as,
N M (F
G
, F
P
) =
,
( ) ( )
( )
G P
G P
H F H F
H F F
(2)
The registration is performed on the transformation space,
T, such that
T = 0
1
0 a b
c d
f e

Where a, b, c, d are shear, scale, and rotation parameters,
and e, f are the translation parameters. Using the normalized
mutual information and exploring the search space, T, we
define a search strategy to find the transformation parameters,
T *.
T * = arg max {N M (F
P,
T (F
G
))} (3)
Gallery and probe face images (F
G
and F
P
) are thus
registered using the transformation T *. This registered
algorithm is linear in nature. To accommodate nonlinear
variations in faces, multi resolution image pyramid scheme is
applied which starts with building the Gaussian pyramid of
both the gallery and probe images. Registration parameters are
estimated at the coarsest level and used to warp the face
images in the next level of the pyramid. The process is
iteratively repeated through each level of the pyramid and a
final transformed gallery face image is obtained at the finest
pyramid level.
In this manner, the global variations at the coarsest
resolution level and local non-linear variations the finest
resolution level are addressed. Once the age difference between
the gallery and probe face image is minimized, face recognition
algorithm can be efficiently applied to verify the identity of the
probe image, By Hazym Kemal Ekenel, Hua GAO[3]proposed
method for facial image preprocessing Discrete Cosine
Transform-Based Local Appearance Model .
Local appearance face recognition is bases on statistical
representations of the non- overlapping local facial regions and
their combination at the feature level. The underlying idea is to
utilize local information while preserving spatial relationships.
In[5], the discrete cosine transform (DCT) is proposed to be
used to represent the local regions. It has been shown to be a
better representation method for modeling the local facial
appearance compared to principal component analysis(PCA)
and the discrete wavelet transform (DWT) in terms of face
recognition performance. Feature extraction from depth images
using local appearance-based face representation can be
summarized as follows:
The input depth image is divided into blocks of 8x8 pixels
size. Each block is then represented by its DCT coefficients.
These DCT coefficients are ordered using the zigzag scanning
pattern [6]. From the ordered coefficients, M of them is
selected according to the feature selection strategy, resulting in
an M-dimensional local feature vector. Finally, the DCT
coefficients extracted from each block are concatenated to
construct the overall feature vector of the corresponding depth
image. In order to compare the introduced local DCT-based
representation with the depth representation, we calculated the
ratio of within class variance with each representation on a
training set which has also been used for identification
experiments.
We calculated the ratio of within class variance to between
class variance for each representation unit and then averaged it
over the representation units and the subjects. We obtained an
average ratio of 0.5 with DCT-based local representation, and
0.67 with the depth representation. The lower ratio of within
class variance to between class variance obtained by the
proposed representation scheme indicates its better
discrimination capability compared to the depth representation
III. A NEW APPROACH FOR FACIAL IMAGE PREPROCESSING
In this section we have discuss the transform of facial
image data and resolve the color of skin into HSV color space
and local binary pattern method. First we have discuss integer
wavelet packet transform function.
A. Integer Wavelet Packets Transform
In this section the wavelet transform and of the filter bank
scheme are given, and the wavelet packets transform are
introduced. The block scheme of the single level wavelet
transform is shown in Fig.1.

Fig. (1) One level transforms function
The low-pass analysis filters and the high-pass ones are
followed by down sampling of a factor two. At the
reconstruction side, the low-pass and band-pass branches are
psampled and filtered with the synthesis filters H(z) and G(z) in
order to obtain the original signal A wavelet transform on J
levels is obtained by iterating the filter bank J-1 times on the
low-pass branch. The wavelet transform coefficients consist of
the J high-pass and the erminal low-pass node sequences output
by the filter bank tree. Given a perfect reconstruction filter
bank, the iterated scheme represents an either orthonormal or
biorthogonal (non-redundant) representation of the original
Vol. 2, No. 11, 2011
146 | P a g e
signal. Differently from the wavelet transform, the J-level WPT
are achieved by iterating the one level filter bank on both the
low-pass and the high-pass branch, and then applying a pruning
algorithm to select a suitable representation. An algorithm has
been proposed in [5], which selects the best representation of a
sequence across the entire tree based on some proper cost
function, which must measure the compactness of the
representation.
B. HSV colour model
The HSV stands for the Hue, Saturation, and Value based
on the artists (Tint, Shade, and Tone). The coordinates system
in a hexacone is shown in Figure 1. (a) And Figure 1. (b) a
view of the HSV color model. The Value represents intensity
of a color, which is decoupled from the color information in the
represented image. The hue and saturation components are
intimately related to the way human eye perceives color
resulting in image processing algorithms with physiological
basis

Fig. (2) HSV color model

C. SVM classifier
Support Vector Machines are based on the concept of
decision planes that define decision boundaries. A decision
plane is one that separates between a set of objects having
different class memberships. A schematic example is shown in
the illustration below. In this example, the objects belong either
to class GREEN or RED.
The separating line defines a boundary on the right side of
which all objects are GREEN and to the left of which all
objects are RED. Any new object (white circle) falling to the
right is labeled, i.e., classified, as GREEN (or classified as
RED should it fall to the left of the separating line).

Fig. (3) Classification of data through SVM
The above is a classic example of a linear classifier, i.e., a
classifier that separates a set of objects into their respective
groups (GREEN and RED in this case) with a line. Most
classification tasks, however, are not that simple, and often
more complex structures are needed in order to make an
optimal separation, i.e., correctly classify new objects (test
cases) on the basis of the examples that are available (train
cases). This situation is depicted in the illustration below.
Compared to the previous schematic, it is clear that a full
separation of the GREEN and RED objects would require a
curve (which is more complex than a line.
Steps for processing of preprocessing of facial image data
(1) Processed scanned image of face into transform
function.
(2) Apply integer wavelet packet transform function for
decomposition of image on the basis of low and high
frequency without loss of data.
(3) Built packet of image data.
(4) Apply HSV color model for better resolution for
facial skin.
(5) Finally used SVM clasifer for classification of low
data and high data
(6) Then apply LBP (local pattern method)

Fig.(4) Flow chart
Vol. 2, No. 11, 2011
147 | P a g e
IV. CONCLUSION AND FUTURE SCOPE
In this paper we discuss a combined pre-processing method
for facial image data for better processing of raw data for
training and detection of face. We have seen here by
mathematics integer wavelet packet transform is a lossless
decomposition technique and very efficient process for image
transformation. Now in future we have implanted this approach
and find some result and these results compare with standard
methods
REFERENCES
[1] Mayank Vatsa, Richa Singh, Samarth Bharadwaj, Himanshu S. Bhatt
and Afzel Noore Matching Digital and Scanned Face Images with Age
Variation IEEE Transactions on Circuits, Systems and Video
Technology, vol. 14, no. 1, pp. 5057, 2010
[2] Liu, C. and Wechsler, H. (2002). Gabor feature based classification
using the enhanced fisher linear discriminant model for face recognition
Image processing, IEEE Transactions on 11:467-476.
[3] Romby, H.A, Baluja, S. and Kanade T. (1998). Neural network based
face detection. Pattern Analysis and Machine Intelligence, IEEE
Transactions on 20:23-38
[4] D. Hill, C. Studholme, and D. Hawkes, Voxel similarity measures for
automated image registration, in Proceedings of Third SPIE Conference
on Visualization in Biomedical Computing, 1994, pp.205-206.
[5] F. Maesa, A. Collignon, D. Vandermeulen, G. Marchal, and P. Suetens,
Multimodality image registration by maximization of mutual
information,IEEE Transactions on Medical Imaging, vol. 6, no.
2,pp.187198, 1997.
[6] H. K. Ekenel and R. Stiefelhagen, Local appearance-based face
recognition using discrete cosine transform, presented at the 13
th
Eur.
Signal Processing Conf., Antalya, Turkey, 2005.
[7] R. C. Gonzales and R. E. Woods, Digital Image Processing. Upper
Saddle River, NJ: Prentice-Hall, 2001
Vol. 2, No. 11, 2011
148 | P a g e
A new approach of designing Multi-Agent Systems
With a practical sample
Sara Maalal
Team of Systems Architecture, Laboratory of computing,
Systems and Renewable Energy
National and High School of Electricity and Mechanic
ENSEM BP 8118, Oasis
Casablanca, Maroc

Malika Addou
Team of Systems Architecture Laboratory of computing,
Systems and Renewable Energy
Hassania School of Public
Works EHTP BP 8108, Oasis
Casablanca, Maroc

AbstractAgent technology is a software paradigm that permits
to implement large and complex distributed applications [1]. In
order to assist analyzing, conception and development or
implementation phases of multi-agent systems, weve tried to
present a practical application of a generic and scalable method
of a MAS with a component-oriented architecture and agent-
based approach that allows MDA to generate source code from a
given model. Weve designed on AUML the class diagrams as a
class meta-model of different agents of a MAS. Then we
generated the source code of the models developed using an open
source tool called AndroMDA. This agent-based and evolutive
approach enhances the modularity and genericity developments
and promotes their reusability in future developments. This
property distinguishes our design methodology of existing
methodologies in that it is constrained by any particular agent-
based model while providing a library of generic models [2].
Keyword- Software agents; Multi-agents Systems (MAS); Analysis;
Software design; Modeling; Models; Diagrams; Architecture;
Model Driven Architecture (MDA); Agent Unified Modeling
Language (AUML); Agent Modeling Language (AML).
I. INTRODUCTION
Currently the computer systems are increasingly complex,
often distributed over several sites and consist of software
interacting with each other or with humans. The need for model
human behavior in specific computer programs has prompted
officials to use technology that affected the last decade and
whose movements are very remarkable. In this context,
designing multi-agent systems (MAS) is complex because they
require the inclusion of several parts of the system which can
often be approached from different angles. We must identify
and analyze all system problems to find models for multi-
agents to implement and integrate them into a coherent system.
This is the software engineering and well justifies the use of a
method of analysis, design and development of multi-agents
systems [2].
This paper describes a practical example of a new generic
model designed for modeling multi-agent systems and based on
a class diagram, defining the different types of agents and
meeting our needs for development and testing of MAS
applications.
II. MULTI-AGENT SYSTEMS
A. Definitions
- An agent is a computer system within an environment and
with an autonomous behavior made for achieving the
objectives that were set during its design [3].
- A multi-agents system is a system that contains a set of
agents that interact with communications protocols and are able
to act on their environment. Different agents have different
spheres of influence, in the sense that they have control (or at
least can influence) on different parts of the environment.
These spheres of influence may overlap in some cases; the fact
that they coincide may cause dependencies reports between
agents [4].
The MAS can be used in several application areas such as
e-commerce, economic systems, distributed information
systems, organizations...
B. Types of agent
Starting from the definitions cited above, we can identify
the following agent types [5]:
The reactive agent is often described as not being
"clever" by itself. It is a very simple component that
perceives the environment and is able to act on it. Its
capacity meets mode only stimulus-action that can be
considered a form of communication.
The cognitive agent is an agent more or less intelligent,
mainly characterized by a symbolic representation of
knowledge and mental concepts. It has a partial
representation of the environment, explicit goals, it is
capable of planning their behavior, remember his past
actions, communicate by sending messages, negotiate,
etc..
The intentional agent or BDI (Belief, Desire and
Intention) is an intelligent agent that applies the model
of human intelligence and human perspective on the
world using mental concepts such as knowledge,
beliefs, intentions, desires, choices, commitments. Its
behavior can be provided by the award of beliefs,
desires and intentions.
Vol. 2, No. 11, 2011
149 | P a g e
The rational agent is an agent that acts in a manner
allowing it to get the most success in achieving the
tasks they were assigned. To this end, we must have
measure of performance, if possible objective
associated with a particular task that the agent should
run.
The adaptive agent is an agent that adapts to any
changes that the environment can have. He is very
intelligent as he is able to change its objectives and its
knowledge base when they change.
The communicative agent is an agent that is used to
communicate information to all around him. This
information can be made of his own perceptions as it
may be transmitted by other agents.

Figure 1. Types of agents
III. THE DESIGN METHODOLOGIES STATE OF THE ART
Building high quality software for real-world applications is
a difficult task because of the large number and the flexibility
of components but also because of the complexity of
interconnections required. The role of software engineering is
precisely that of providing methodologies that can facilitate
control of this complexity. A methodology by definition can
facilitate the process of engineering systems. It consists of
guides that cover the entire lifecycle of software development.
Some are technical guides; others are managing the project [6].
Well name method the approach to use a rigorous
process for generating a set of models that describe various
aspects of software being developed using a well- defined
notation.
To this end, several software engineering paradigms have
been proposed, such as object-oriented design patterns, various
software architectures. These paradigms fail especially when it
concerns the development of complex distributed systems for
two reasons: the interactions between the various entities are
defined in a too rigid way and there is no mechanism complex
enough to represent the organizational structure system [7]. The
paradigm of agents and multi-agent systems can be a good
answer to these problems, because the agent-oriented
approaches significantly increase our ability to model, design
and build complex distributed systems [8].
There are many methodologies for analysis and design of
multi-agent systems. We cite below some examples of existing
methodologies [2]:
The AAII methodology was developed based on the
experience accumulated during the construction of BDI
systems. In this methodology, we have a set of
templates that, when they have been fully elaborated,
define the specifications of agents such as desires,
beliefs and intentions [9].
The first version of Gaia methodology, which modeled
agents from the object-oriented point of view, was
revisited 3 years later by the same authors in order to
represent a MAS as an organized society of individuals
[10]. In fact, the agent entity, which is a central
element of the meta-model of Gaia, can play one or
more roles. A role is a specific behavior to be played
by an agent (or kind of agents), defined in term of
permissions, responsibilities, activities, and interactions
with other roles. When playing a role, an agent updates
its behavior in terms of services that can be activated
according to some specific pre- and post- conditions. In
addition, a role is decomposed in several protocols
when agents need to communicate some data. The
environment abstraction specifies all the entities and
resources a multi-agent system may interact with,
restricting the interactions by means of the permitted
actions [1].
The Gaia methodology gives the possibility to design
MAS using an organizational paradigm and to traverse
systematically the path that begins by setting out the
demands of the problem and to lead to a fairly detailed
and immediate implementation [9]. Gaia permits to
design a hierarchical non-overlapping structure of
agents with a limited depth. From the organizational
point of view, agents form teams as they belong to a
unique organization, they can explicitly communicate
with other agents within the same organization by
means of collaborations, and organizations can
communicate between them by means of interactions.
If inter-organization communication is omitted,
coalitions and congregations may also be modeled [1].
However, this methodology is somewhat limited since
we can describe MAS with different architectures of
agents [9].
The main contribution of MESSAGE was the
definition of meta-models for specification of the
elements that can be used to describe each of the
aspects that constitute a multi-agent system (MAS)
from five viewpoints: organization, agents, goals/tasks,
interactions and domain. MESSAGE adopted the
Unified Process and centered on analysis and design
phases of development [11].
INGENIAS starts from the results of MESSAGE and
provides a notation to guide the development process
of a MAS from analysis to implementation [12] [13].
It is both a methodology and a set of tools for
development of multi-agent systems (MAS). As a
methodology, it tries to integrate results from other
proposals and considers the MAS from five
complementary viewpoints: organization, agent,
Vol. 2, No. 11, 2011
150 | P a g e
tasks/goals, interactions, and environment. It is
supported by a set of tools for modeling (graphical
editor), documentation and code generation (for
different agent platforms). The INGENIAS
methodology does not explicitly model social norms,
although they are implicit in the organizational
viewpoint. Organizational dynamics are not considered
i.e., how agents can join or leave the system, how they
can form groups dynamically, what their life-cycle is,
etc [14]. The authors have developed an agent-oriented
software tool called INGENIAS Development Kit
(IDK) [15]. It allows to edit consistent models
(according to INGENIAS specification) and to
generate documented code in different languages such
as JADE [16], Robocode, Servlets or Gracias Agents
[1].
Multi-agent systems Software Engineering (MaSE) is a
start-to-end methodology that covers from the analysis
to the implementation of a MAS [17]. The main goal of
MaSE is to guide a designer through the software life-
cycle from a documented specication to an
implemented agent system, with no dependency of a
particular MAS architecture, agent architecture,
programming language, or message-passing system.
AUML (Agent Unified Modeling Language) is an
evolving standard for a design methodology to support
MAS. It is based on the UML methodology used with
object oriented systems. This notation was proposed to
adapt the UMLs one in order to describe the agent-
oriented modeling [18].

AUML provides tools for:
Specification protocol of interaction between
agents,
Representation of the internal behaviour of an
agent,
Specification of roles, package interface agent,
mobility, etc [2].

The Agent Modeling Language (AML) is a semiformal
visual modeling language for specifying, modeling and
documenting systems that incorporate concepts drawn
from multi-agents systems (MAS) theory [19].
ASPECS (Agent-oriented Software Process for
Engineering Complex Systems) provides a holonic
perspective to design MAS [20]. Considering that
complex systems typically exhibit a hierarchical
configuration, on the contrary to other methodologies,
it uses holons instead of atomic entities. Holons, which
are agents recursively composed by other agents,
permit to design systems with different granularities
until the requested tasks are manageable by individual
entities.
The goal of the proposed meta-model of ASPECS is to
gather the advantages of organizational approaches as
well as of those of the holonic vision in the modeling
of complex system [1].
All these methodologies presented above are still quite
recent. They are mainly focused on the analysis phase, whereas
design and implementation phases are missing or are redirected
to agent-oriented methodologies, which do not oer enough
tools to model organizational concepts. Therefore, there is still
a gap between analysis and design, which must be specied
clearly, correctly and completely [14].
Finally, the maturity of methodologies can be analyzed by
the number of systems that have adopted them. Most of
analyzed methodologies have associated applications that show
their feasibility. These methodologies have been applied in
different elds such as medical informatics [21], manufacturing
[20] [22], and e-commerce [23]. MaSE and INGENIAS are the
most used ones. Unfortunately, the number of real world
applications that use agent-oriented methodologies is still low
[1].
IV. THE MDA APPROACH
The MDA (Model Driven Architecture) proposes a
methodological framework and architecture for systems
development that focuses first on the functionality and
application behavior, without worrying about the technology
with which the application will be implemented. The
implementation of the application goes through the
transformation of business models in specific models to a target
platform (Fig.2). One research was done in this area as the
dissertation of Jarraya T. [24]

Figure 2. The MDA approach
The business process independent of automation,
which comes from the expression of need, is
described as a "CIM" (Computation Independent
Model). The detailed functional analysis, the heart of
the process is concentrated in the "PIM" (Platform
Independent Model), which, as its name suggests, is strictly
independent of the technical architecture and the target
language. The "PSM" (Platform Specific Model) is the
model for engineering design obtained by transformation
of PIM by projection on the target technical architecture. It is
this model that is based on code generation [5].
The benefits to businesses on the MDA are primarily:
Vol. 2, No. 11, 2011
151 | P a g e
The fact that architectures based on MDA are ready
for technological developments.
The ease of integrating applications and systems
around a shared architecture
Broader interoperability for not being tied to a
platform.
One of the main tools of MDA, we have AndroMDA who
takes as its input a business model specified in the Unified
Modeling Language (UML) and generates significant portions
of the layers needed to build, for example, a Java application
[25]. AndroMDA's ability to automatically translate high-level
business specifications into production quality code results in
significant time savings when implementing Java applications.
The diagram below maps various application layers to, for
examples, Java technologies supported by AndroMDA [5].

Figure 3. Application layers supported by AndroMDA
Presentation Layer: AndroMDA currently offers two
technology options to build web based presentation
layers: Struts and JSF. It accepts UML activity
diagrams as input to specify page flows and generates
Web components that conform to the Struts or JSF
frameworks.
Business Layer: The business layer generated by
AndroMDA consists primarily of services that are
configured using the Spring Framework. These
services are implemented manually in AndroMDA-
generated blank methods, where business logic can be
defined. These generated services can optionally be
front-ended with EJBs, in which case the services
must be deployed in an EJB container (e.g.,JBoss).
Services can also be exposed as Web Services,
providing a platform independent way for clients to
access their functionality. AndroMDA can even
generate business processes and workflows for the
jBPM workflow engine (part of the JBoss product
line).
Data Access Layer: AndroMDA leverages the
popular object-relational mapping tool
called Hibernate to generate the data access layer for
applications. AndroMDA does this by generating Data
Access Objects (DAOs) for entities defined in the
UML model. These data access objects use the
Hibernate API to convert database records into objects
and vice-versa. AndroMDA also supports Enterprise
Java Beans EJB3/Seam [26] for data access layer (pre-
release).
Data Stores: Since AndroMDA generated
applications use Hibernate to access the data, you can
use any of the databases supported by Hibernate.
The generation process of AndroMDA is as follows [5] :

Figure 4. Generation process of AndroMDA
Preparation of the project in MagicDraw
Preparing use cases
Preparation of class diagram
Preparation of state charts
Code Generation
Generating the database
Deploy the application

V. PROPOSED APPROACH

Our approach is based on model driven architecture (MDA)
which aims to establish the link between the existing agent
architectures and models or meta-model multi-agent systems
that we build based on AUML. Our idea is to offer a design
methodology based on agents AUML notation for establishing
a generic class diagram that the designer can use to design his
system [3]. This diagram is considered as a meta-model which
Vol. 2, No. 11, 2011
152 | P a g e
is not generated by any tool and must be defined by the
modeler himself.

Figure 5. An AUML generic class diagram for a MAS
Our approach has a lot of benefits, it allows:
Reducing costs and development times for new
applications.
Improving quality of applications.
Reducing complexity of application development.
Ability to generate all the necessary components
described.
Modularity and reusability of the developments.
Coercion by the MDA model.
Generating a library of generic models.
A. Description of the AUML generic Class Diagram
The diagram is conceived in three layers, each one is
represented by a relationship between classes: A first part
which is a relation between agent and its environment, a second
part of specialisation of the agent class, and at the last part, a
specialisation of the cognitive agent class [3].
1- The first part
The first part consists of two important classes:
- Environment,
- Agent

Environment is an important class on the diagram
because it influences all the system. Environments
data is represented by two sections, Attributes and
Perceptions. Attributes can be all the information that
an environment should have, plus the following
common information:

Deterministic when the next state of the
environment is determined in a unique way
by the current state and action of the agent,
so the environment is deterministic. If the
outcome is uncertain (especially if, as a result
of action of the agent, the environment can
evolve in different ways), we are in the non-
deterministic case.
Static if the environment cannot change its
state without the intervention of the agent.
The environment is dynamic if its state can
change without the action of the agent in the
time interval between two perceptions of the
agent.
Continuous if any portion of an environment
state to another requires passing through a
sequence of intermediate states, otherwise
the environment is discrete.
Perception is a section where the designer should
determinate all environment perceptions, example:
number of agents.

Environment contains several functions allowing to
start running, to perceive information from agents
linked to it and to modify its state after each action
from those agents, that is respectively Run(),
Perceive() and ModifState().

Agent is the main class on the diagram that allows the
designer to express all agent properties. The
constructor of Agents takes three sections: Roles,
Attributes and Perception. Roles are agent
functionalities. Attributes are all information that an
agent should possess. And finally Perception which is
a section where the designer should determinate all
agents perceptions about his environment or the other
agents.

Agent contains several functions who allows starting
running and perceiving information from environment
or agents linked to it and to execute all its actions, that
is respectively Run(), Perceive() and Act().
The first part consists also of two important association
classes:
-Action, between agent and his environment.
-Interaction, between agents.

Action is an association class between agent and
environment. It lists all possible actions that an agent
can execute on his environment.

Interaction is a reflexive association class between
agents. Agent can request information by the
getInformation() function and send it by the inform()
function. Agent may also deal with some constraints
Vol. 2, No. 11, 2011
153 | P a g e
that it is possible to inform by the function
informaboutConstraintes(). The acceptance of
partnership is added also to the main functionalities
of Agent by the function acceptPartnerShip().
2- The second part
The second part represents a specialisation relation of the
Agent class. It consists of three important classes:
- Reactive agent,
- Cognitive agent,
- Communicative agent.

Reactive agent is a type of agent. It possesses the
same properties of the Agent class.

Cognitive agent is another specialization of the Agent
class. In this class, the designer should determinate the
representations of the agent that he must have during
its execution. The class possesses also one important
function Decide() where agent can decide to execute
an action or not according to his goals.

Communicative agent is the last specialization of the
Agent class. Like Cognitive agent class,
Communicative agent class has representations but
possesses a different function called Communicate()
where agent must use to communicate his information
to the other agents.
3- The third part
The third part represents a specialization relation of the
Cognitive agent class. It consists of three important classes:
- Adaptive agent,
- Intentional agent,
- Rational agent.

Adaptive agent is a type of cognitive agent. It
possesses the same properties of the Agent class, the
knowledge base and the Decide() function. As
mentioned in the types of agent section above, an
adaptive agent is able to change its objectives and its
knowledge base as and when these changes. This
functionality is expressed by the
Change_information() function.
Intentional agent or BDI Agent is designed from the
"Belief-Desire-Intention model. It is a type of
cognitive agent. In the same case of Adaptive Agent
class, this class possesses the same properties of the
Agent class, the knowledge base and the Decide()
function.
In this class, the designer should determinate the
agents beliefs represented by the Beliefs section. The
beliefs of an agent are the information that the agent
has on the environment and other agents that exist in
the same environment. Beliefs may be incorrect,
incomplete or uncertain, and because of that, they are
different from knowledge of the agent, which is
information still true. Beliefs can change over time as
the agent by its ability to perceive or interact with
other agents, collects more information.
The designer should also determinate the agents
intentions represented by the Intentions section. The
intentions of an agent are the actions it has decided to
do to accomplish his goals.
To choose the correct agents beliefs from the
incorrect ones, this class offers the
Revise_beliefs(Pres, Belief) function which is based
on the agents knowledge base and his beliefs. Then,
the Generate_desires(Belief, int) function comes to
generate all the agents desires that he may be able to
accomplish at once. The desires of an agent
representing all things the agent would like to see
made. An agent may have conflicting desires, in
which case he must choose between her desires a
subset that is consistent. This subset consists of his
desires is identified with the beliefs and the intentions
of the agent.
Another function comes after that, the Filter(Belief,
Generate_desires, int) which filters all those elements
above and gives the consistent beliefs, desires and
intentions of the intentional agent.
Finally, the agent can select his actions according to
this filtration and execute them by the
Actions_selection(Filter) function.

Rational agent is the last specialisation of the
Cognitive Agent class. Like Intentional Agent class,
Rational Agent class has the Beliefs and the Intentions
sections but possesses just one function called
Mesure_performance(Percept, Belief) where agent
must use to execute his actions as efficient as possible.
This function is based both on his perceptions and his
beliefs.
B. The generic UML Class Diagram
This generic AUML class diagram was subsequently
converted into a generic class diagram based on UML notation.
This transformation will allow the designer to easily use
AndroMDA to generate the source code equivalent to its UML
diagram [1].
The passage from AUML to UML was performed by
following the steps below:

1. Keep the same titles of classes and associations
which constitute the AUML diagram.
2. Assign roles, perceptions, intentions, beliefs and
representations of each agent, and any possible
additional attributes, in the attributes part of the
UML class.
3. Combine all methods or functions in the
operations part of the UML class.
We can obtain, in the end, the following result shown in
Fig. 6:

Vol. 2, No. 11, 2011
154 | P a g e

Figure 6. An UML generic class diagram for a MAS

Our approach can present one desadvantage. It is the
complexity of generating a good code source by AndroMDA.
The model developed at the design phase, should be reliable in
order to build the application and realize its implementation
without errors [5].
V. APPLICATION EXAMPLE
A. Description
Our proposed AUML class diagram was used for design of
one multi-agent system for a Chat Application. This example is
designed as follows [5]:
Three reactive agents: These agents will be the chatters,
the interest that these are reactive agents relies on the fact
that an agent doesn't react before the declaration of the
name of the receiver by the user of the application.
Therefore an agent will react to get ready to catch the
name and the message and to send it to the appropriate
person. He will react also to clear the sent and the received
message from their area in his interface.

We can respectively obtain the following AUML and UML
diagrams corresponding to this example, shown in the Figures
7 and 8:
B. Realization
To validate our model for this example, weve tried to
download AndroMDA with all the required dependencies
(including all profiles referenced by models). Then, we
generated our project ChatAgents by running mvn
org.andromda.maven.plugins:andromdaapp-maven
plugin:3.4-SNAPSHOT:generate . The result of this
command is as follows:

Figure 7. AUML Class diagram for a chat application

Figure 8. UML Class diagram for a chat application
When we examine the various folders and files created by
the andromdapp plug-in, we will notice files called pom.xml in
various folders under ChatAgents. These files make up several
Maven projects. In fact, the ChatAgents directory contains a
hierarchy of Maven projects as shown below [5].
ChatAgents: This is the master project that controls
the overall build process and common properties.
mda: The mda project is the most important sub-
project of the application. It houses the ChatAgents
UML model under the src/main/uml directory. The
mda project is also where AndroMDA is configured
to generate the files needed to assemble the
application.
common: The common sub-project collects resources
and classes that are shared among other sub-projects.
These include value objects and embedded values.

Vol. 2, No. 11, 2011
155 | P a g e

Figure 9 : ChatAgents project generation

ChatAgents
|
|-- mda
|
|-- common
|
|-- core
|
|-- web
|
+-- app
core: The core sub-project collects resources and
classes that use the Spring framework, optionally
making use of Hibernate and/or EJBs under the hood.
These include entity classes, data access objects,
hibernate mapping files, and services.
web: The web sub-project collects those resources
and classes that make up the presentation layer.
app: The app sub-project collects those resources and
classes that are required to build the .ear bundle.
By opening the file ChatAgents.xml in MagicDraw, we
will be able to build various graphs of our model to
generate then the source code of the entire
application. Note that AndroMDA can't read MagicDraw 17
models directly. Therefore, you can export it to another file
format: EMF-UML2.
After import of AndroMDA profiles to use for our
application, we designed our class diagram as shown in Fig.10
as follows [5]:
The result of exporting our ChatAgents model to EMF-
UML2 format is located in the
folder C:/ChatAgents/mda/src/main/uml in explorer. Below his
content:
ChatAgents.xml: the MagicDraw 17 model file.
ChatAgents.uml: ChatAgents model in EMF/UML2
format. It's the file that will be processed by
AndroMDA.
10 files ending with .profile.uml: the different
profiles used by ChatAgents.uml
Following the definition of our model, the generation
of application code is achieved by executing the command
"mvn install", the result appears as in the figure [5].
Thus, the class Chat.java is created and can be
easily accessed and modified by the developer where he has the
ability to implement its operations in the generated code.
We conducted this implementation and got the final result.
VI. CONCLUSION AND FUTURE SCOPE
The purpose of this paper is to demonstrate the feasibility of
our approach to analyze, design and implement multi-agent
systems. With AUML modeling and MDA, we can generate all
the necessary components described by the class meta-model
that we proposed. Which leads us to obtain a generic design
based on SOA more or less reusable components using one of
the most MDA tools used in development is AndroMDA [27].
In the future, we would like to model another application
sample of our model but in a more complex form using
cognitive or adaptive agents and in other platforms like C++,
Web services, etc. It will help us to validate the efficacy of our
proposed approach and lead us to consider it as a generic
approach which can be adopted by every type of information
system and used for any real world application.

Vol. 2, No. 11, 2011
156 | P a g e
Figure 10 : Class diagram built on MagicDraw 17

Figure 11 : Code generation after definition model

Figure 12. Chat application with three agents
ACKNOWLEDGMENT
I would like to thank to my advisor Ms. M. Addou, Phd. for
his invaluable guidance and many useful suggestions during
my work on this paper. I would also like to express my
gratitude to all those who gave me the possibility to complete
this paper.

Vol. 2, No. 11, 2011
157 | P a g e
REFERENCES
[1] D. Isern, D.Sanchez, A.Moreno, Organizational structures supported by
agent-oriented methodologies, The journal of Systems and Software,
vol. 84, n. 2, Oxford, UK: Elsevier, 2011, pp. 169-184.
[2] S. Maalal, M. Addou, A Model Design of Multi-Agent Systems,
Proceedings of the 2nd Edition of the IEEE International Conference on
Multimedia Computing and Systems ICMCS11, Ouarzazate Morocco,
p. 674, 2011.
[3] M. Wooldridge, Intelligent Agents, Multi agent systems, In The MIT
Press, A modern Approach to Distributed Artificial Intelligence,
(England Massachutts London: MIT Press Cambridge, 1995, p. 27-78)
[4] M. Wooldridge, An Introduction to Multi-Agent Systems, Wiley &
Sons, 2000.
[5] S. Maalal, M. Addou, A practical application of a method of designing
multi-agent systems based on the AUML language and the MDA
approach, Proceedings of the Fourth Workshop on Information
Technologies and Communication WOTIC11, Casablanca, Morocco,
p.104, 2011.
[6] O. Shehory, A. Sturm, Evaluation of modeling techniques for agent-
bases systems, Proceedings of the 5th International Conference on
Autonomous Agents, pp.624-631, 2001.
[7] N. R. Jennings, On agent-based software engineering, Artificial
Intelligence, vol. 117, pp. 277-296, 2000.
[8] M. Wooldridge, N. R. Jennings, Intelligent agent: Theory and practice,
The Knowledge Engineering Review, Vol. 10, n. 2, pp. 115-152, 1995.
[9] A. M. Florea, D. Kayser, S. Pentiuc, A. El Fallah Segrounichi,
Intelligents agents, Agents Intelligents, Politechnica University of
Bucharest, 2002.
[10] L. Cernuzzi, T. Juan, L.Sterling, F. Zambonelli, The Gaia methodology:
basic concepts and extensions, Methodologies and Software
Engeneering for Agent Systems, US: Springer, pp.69-88, 2004.
[11] J. Pavn, , J. Gmez-Sanz., Agent Oriented Software Engineering with
INGENIAS, Proceedings of the international Central and Eastern
European conference on Multi-Agent Systems CEEMAS03, pp.394-
403, 2003.
[12] R. Fuentes-Fernndez, I. Garca-Magariio, A.M. Gmez-Rodrguez,
J.C. Gonzlez-Moreno, A technique for defining agent-oriented
engineering processes with tool support, Artificial Intelligence, vol.23,
pp.432-444.
[13] J. Pavn, , J.J. Gmez-Sanz., R. Fuentes, The INGENIAS methodology
and tools in Agent-oriented Methodologies, B. Henderson-Sellers and
P. Giorgini Eds. Idea Group, 2005, pp. 236276.
[14] E. Argente, V. Julian, V. Botti, Multi-agent system development based
on organizations, Electronic Notes in Theoretical Computer Science,
vol.150, pp.55-71, 2006.
[15] IDK (INGENIAS Development Kit),
http://sourceforge.net/projects/ingenias/
[16] JADE (Java Agent DEvelopment Framework), http://jade.tilab.com/.
[17] S.A. DeLoach, The MaSE methodology, in Methodologies and
Software Engineering for Agent Systems, F. Bergenti, M.P Gleizes, F.
Zambonelli, Eds. The Agent-oriented Software Engineering Handbook.
Kluwer Academic Publishers, 2004, pp. 107125
[18] S. Lynch, K. Rajendran, Design Diagrams for Multi-agents Systems,
Proceedings of the 16th Annual Workshop of the Psychology of
Programming Interest Group PPIG04, pp. 66-78, 2004.
[19] R. Cervenka, I. Trencansky, Agent Modeling Language (AML): A
Comprehensive Approach to Modeling MAS, Informatica, vol. 29, n. 4,
pp. 391-400, 2005.
[20] M. Cossentino, N. Gaud, V. Hilaire, S.Galland, A. Koukam, ASPECS:
An Agent-oriented Software Process for Engineering Complex Systems:
How to design agent societies under a holonic perspective, 2010.
[21] D. Isern, C. Gmez-Alonso, A. Moreno, Methodological development
of a multi-agent system in the healthcare domain, Commun, SIWN 3,
pp. 6568, 2008.
[22] A. Giret, V. Botti, S. Valero, MAS methodology for HMS, In the
Second International Conference on Industrial Applications of Holonic
and Multi-Agent Systems HoloMAS, Springer-Verlag, Copenhagen,
Denmark, pp. 3949, 2005.
[23] J. Ferber, O. Gutknecht, F. Michel, From agents to organizations: an
organizational view of multi-agent systems, in Springer-Verlag Berlin
Heidelberg, P. Giorgini, J. Mller, J. Odell, Eds 2003, in the 4th
International Workshop on Agent-oriented Software Engineering IV
(AOSE), Melbourne, Australia, pp. 214230, 2003.
[24] T. Jarraya, Re-use of interaction protocols and Career-oriented models
for multi-agents development, Rutilisation des protocoles dinteraction
et Dmarche oriente modles pour le dveloppement multi-agents ,
Ph.D. Thesis, Dept. Computer Engineering, University of Reims
Champagne Ardenne, France, 2006.
[25] N. Bhatia, Getting Started with AndroMDA for Java
(www.andromda.org, 2010).
[26] JBoss Seam (http://www.jboss.com/products/seam/).
[27] S. Maalal, M. Addou, A Model Design of Multi-Agents Systems, in
the International Conference on Models of Information and
Communication Systems MICS10, Rabat, Morocco, 2010, unpublished.
AUTHORS PROFILE
Sara Maalal was born in Rabat the Moroccos capital in 1985. She received
his professional master in Computer Engineering and Internet (3I), Option:
Security Networks and Systems, in 2008 from the Faculty of science of
HASSAN II University, Casablanca, Morocco. In 2010 she joined the system
architecture team of the National and High School of Electricity and Mechanic
(ENSEM: Ecole Nationale Suprieure dElectricit et de Mcanique),
Casablanca, Morocco.
Her actual main research interests concern Designing and modeling Multi-
Agent Systems.
Ms. Maalal is actually a Software Engineer in a Moroccan multinational
society called Hightech Payment Systems (HPS) which has always proved
itself as a leading payment solutions provider.

Malika Addou received her Ph.D. in Artificial Intelligence from University of
Liege, Liege, Belgium, in 1992. She got her engineer degree in Computer
Systems from the Mohammadia School of Engineers (EMI : Ecole
Mohammadia des ingnieurs), Rabat, Morocco in 1982. She is Professor of
Computer Science at the Hassania School of Public Works (EHTP : Ecole
Hassania des Travaux Publics), Casablanca, since 1982.
Her research focuses on Software Engineering (methods and technologies for
design and development), on Information Systems (Distributed Systems) and
on Artificial Intelligence (especially Multi-Agent Systems technologies).

IJACSA Volume2No11

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

IJACSA Volume2No11

Uploaded by

Copyright:

Available Formats

(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 2, No. 11, 2011

per time unit. It

the probability of being in state . In a

for any (by using

so that the above probability

the above derivation and result becomes

the size of the

oraring rom to gets elaye

time units to entirely empty the key-

, each of is connected by two directed

denotes the incoming traffic flow

bits of fresh key-material to

shown in Table 1(a) for the incident of

for each link).

} be the data block of bits which Alice is going to

in classic memory and

and the message decoded as

is incorrect then the key

denotes Eve's probe states and is the initial

. But in that case Eve's

sequences and the rest part of these sequences remains

malicious Eve has to

. The number of test qubits must provide

with encoded message is sent. As a

and each block is processed and sent independently

. The block with encoded preprocessing key

is sent as the first one.

(IJACSA) International Journal of Advanced Computer Science and Applications,

You might also like